Agricultural spraying optimization method based on laser radar obstacle avoidance

Abstract

The invention relates to the technical field of agricultural intelligent pesticide application, and discloses an agricultural spraying optimization method based on laser radar obstacle avoidance, and the method comprises the steps: constructing a four-dimensional environment tensor through multi-modal data fusion, planning a collision-free path guided by a dynamic situation field, and matching a crop normal vector, three-dimensional directional spraying is realized by combining unsteady state fluid mechanics and static parameter dynamic regulation and control, and a closed-loop optimization path and spraying parameters are fed back based on laser radar reflection intensity; the system comprises a multi-modal data acquisition module, a four-dimensional environment modeling module, a dynamic path planning module, a directional spraying control module, a closed-loop feedback module and an execution mechanism module. The sensing precision is improved through multi-modal data fusion and four-dimensional modeling, liquid medicine targeted deposition and parameter self-adaptive regulation and control are achieved in combination with dynamic situation field path planning, fluid-electrostatic field coupling and closed-loop feedback optimization, and real-time accurate pesticide application in a complex farmland environment is guaranteed by means of a lightweight edge computing architecture.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural spraying technology, specifically to an agricultural spraying optimization method based on lidar obstacle avoidance. Background Technology

[0002] Traditional agricultural spraying machinery relies heavily on preset paths and fixed spraying parameters, making it difficult to adapt to the dynamic distribution of obstacles, the changing geometry of crops, and the fluctuations in micro-meteorological conditions in complex farmland environments. Existing environmental perception technologies often use single sensors for local obstacle detection, lacking the ability to fuse multi-dimensional data in a spatiotemporal manner. This results in insufficient modeling resolution, especially in densely vegetated areas where blind spots can easily occur due to occlusion, leading to collision risks or overly conservative obstacle avoidance path planning.

[0003] At the path planning level, although mainstream algorithms can generate collision-free trajectories in static environments, their dynamic response speed and replanning efficiency are low when faced with sudden moving obstacles or real-time weather disturbances in farmland scenarios. Furthermore, they do not fully consider the matching requirements between the spray trajectory and the geometric features of the crop surface, which can easily cause non-targeted dispersion of pesticides and reduce the effective deposition rate.

[0004] In terms of pesticide diffusion control, traditional electrostatic spraying technology often relies on empirical formulas to set constant voltage parameters, making it difficult to adapt to dynamic changes in variables such as target distance and ambient wind speed. This results in poor uniformity of droplet deposition after being disturbed by airflow, leading to problems such as excessive pesticide residue in some areas and insufficient coverage in others. In addition, existing systems generally lack real-time monitoring and closed-loop feedback mechanisms for operational effectiveness, and parameter adjustments depend on manual experience, making it difficult to achieve precise matching of pesticide dosage with crop needs.

[0005] Therefore, this invention proposes an agricultural spraying optimization method based on lidar obstacle avoidance to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an agricultural spraying optimization method based on lidar obstacle avoidance, which solves the problems of insufficient perception capability, low dynamic obstacle avoidance efficiency, and poor uniformity of pesticide deposition in complex farmland environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an agricultural spraying optimization method based on lidar obstacle avoidance, comprising the following steps:

[0008] S1. Acquire three-dimensional point cloud data of farmland through lidar, and simultaneously collect thermal imaging data, attitude data and meteorological data to obtain multimodal environmental data;

[0009] S2. Based on the multimodal environmental data, construct a four-dimensional environmental tensor that integrates spatial coordinates, time series, laser reflection intensity, and thermal imaging temperature. Reduce the dimensionality of the four-dimensional environmental tensor through tensor decomposition and fill in the missing regional data caused by occlusion to generate a passable region model.

[0010] S3. Based on the accessible area model, a collision-free path is planned using a dynamic potential field and path search algorithm. The collision-free path includes a spray trajectory that matches the normal vector of the crop surface.

[0011] S4. Based on the collision-free path and real-time environmental parameters, combined with the drug diffusion model constructed based on the unsteady fluid dynamics equation and the electrostatic adsorption control parameters dynamically adjusted based on the three-dimensional point cloud data, three-dimensional directional spraying is performed. The electrostatic adsorption control parameters include a pulsed high voltage dynamically adjusted according to the target distance and wind speed.

[0012] S5. Calculate the amount of drug deposition in real time by using the change rate of lidar reflection intensity to obtain the deposition results, and dynamically adjust the path planning and spray parameters based on the deposition results.

[0013] Preferably, step S1 includes the following steps:

[0014] A 3D point cloud dataset is generated by scanning the farmland environment with LiDAR. , where each point Includes spatial coordinates and timestamp ;

[0015] For each point in the three-dimensional point cloud data Select its local neighborhood point set Calculate the covariance matrix ,in, The neighborhood centroid; with respect to the covariance moments Perform eigenvalue decomposition, and take the eigenvector corresponding to the smallest eigenvalue as a point. normal vector ;

[0016] Simultaneous acquisition of temperature distribution data output by thermal imager IMU-measured device attitude angle Wind speed collected by meteorological sensors and humidity This data is integrated into multimodal environmental data.

[0017] Preferably, step S2 includes the following steps:

[0018] Spatial coordinates in the three-dimensional point cloud data timestamp Reflection intensity and thermal imaging temperature Integrate into a four-dimensional environment tensor according to a preset time window. ,in, For spatial coordinate dimensions; For time series dimensions; For the dimension of reflection intensity, Temperature dimension;

[0019] The four-dimensional environment tensor is subjected to Tucker decomposition to obtain the core tensor. sum factor matrix ,satisfy .

[0020] Preferably, in step S2, the process of completing the missing data due to occlusion includes:

[0021] Based on the core tensor sum factor matrix Define the missing region index set ,in, Indexed by spatial coordinate dimensions; Indexed for time series dimensions; Indexed by the dimension of reflection intensity; Indexing the temperature dimension; constructing an optimization problem:

[0022] ;

[0023] in, For the tensor nuclear norm, The regularization coefficient is used; the optimization problem is solved iteratively using the alternating direction multiplier method, and the completed environment tensor is output. .

[0024] Preferably, step S3, based on the passable area model, uses a dynamic potential field and path search algorithm to plan a collision-free path, including:

[0025] Constructing the dynamic potential function Among them, the gravitational field repulsive field satisfy:

[0026] ;

[0027] in, The current coordinates; The coordinates of the target point; The coordinates of the obstacle; The distance is Euclidean. This is the gain coefficient; This represents the maximum distance at which the obstacle can have an effect.

[0028] Based on the improved RRT* algorithm, an initial path node set is generated within a GPU parallel computing framework. and edge set ;

[0029] During the node expansion phase, based on the potential field gradient direction... Bias the random sampling points to generate new nodes. ; through collision detection function Verify node security, among which, The completed environment tensor; when a dynamic obstacle position update is detected, local path replanning is triggered with a response period synchronized with the multimodal environment data acquisition.

[0030] Preferably, in step S3, the collision-free path comprising a spray trajectory matching the crop surface normal is achieved through the following steps:

[0031] The normal vector extracted from the 3D point cloud data Calculate the spray direction vector With normal vector The included angle According to the included angle Adjust the pitch and yaw angles of the spray nozzles to make... ,in, This is a preset angle threshold.

[0032] Preferably, in step S4, the implementation of the drug diffusion model based on the unsteady hydrodynamic equations includes:

[0033] Establish the governing equations that couple the Navier-Stokes equations with the electrostatic field:

[0034] ;

[0035] in, The density of the liquid medicine; For the velocity field; For pressure; Dynamic viscosity; The electrostatic field strength; It is the dielectric constant; The charge density is used; the governing equations are discretized using the finite volume method, combined with real-time environmental parameters such as wind speed. With humidity Iteratively solve the three-dimensional drug concentration distribution .

[0036] Preferably, in step S4, the dynamic adjustment of the electrostatic adsorption control parameters includes:

[0037] Based on target distance With real-time wind speed Calculate the pulsed high voltage amplitude using the following formula. :

[0038] ;

[0039] in, This is an environmental correction factor; The surface charge of the target crop; the normal vector extracted from the three-dimensional point cloud data. Calculate the direction vector of each ring electrode. With normal vector The included angle ,when The corresponding electrode is activated to generate a directional electrostatic field.

[0040] Preferably, step S5 includes the following steps:

[0041] Acquire lidar reflection intensity data before and after spraying. and Calculate the rate of change of reflection intensity ;

[0042] The rate of change of reflection intensity was established through calibration experiments. With the amount of drug solution deposited mapping relationship ,in, These are calibration coefficients;

[0043] The amount of drug solution deposited is input into a multi-objective reinforcement learning model to minimize the drug solution waste rate. and maximize coverage To optimize the objective, output the adjusted path planning parameters. and spray parameters The reward function of the reinforcement learning model is:

[0044] ;

[0045] in, These are the weighting coefficients.

[0046] The present invention also provides an agricultural spraying optimization system based on lidar obstacle avoidance, the system comprising:

[0047] The multimodal data acquisition module integrates lidar, thermal imager, IMU and meteorological sensors to simultaneously acquire 3D point cloud data, temperature distribution, equipment attitude and environmental parameters;

[0048] The four-dimensional environment modeling module is communicatively connected to the multimodal data acquisition module and is used to construct and complete the four-dimensional environment tensor to generate a passable area model.

[0049] The dynamic path planning module, based on the improved RRT* algorithm and GPU parallel computing unit, is used to plan collision-free paths in real time.

[0050] The directional spray control module includes an unsteady hydrodynamic solver, a ring electrode array, and a high-voltage pulse generator, which are used to perform three-dimensional drug diffusion simulation and dynamic control of electrostatic adsorption parameters.

[0051] The closed-loop feedback module, through the lidar reflection intensity analysis unit and the reinforcement learning decision unit, is used to calculate the amount of drug deposition and optimize the path and spray parameters;

[0052] The actuator module, which includes a multi-degree-of-freedom robotic arm and a fan-shaped nozzle, is used to perform directional spraying operations according to adjusted parameters.

[0053] This invention provides an agricultural spraying optimization method based on lidar obstacle avoidance. It has the following beneficial effects:

[0054] 1. This invention integrates data from lidar, thermal imaging, attitude and meteorological sensors to construct a four-dimensional environmental tensor that integrates spatiotemporal, reflection intensity and temperature. This overcomes the perception limitations of traditional single sensors in complex farmland scenarios, and enables accurate modeling of obstacle boundaries, crop geometry and microenvironment parameters. It provides high-resolution environmental priors for dynamic path planning and directional spraying, and significantly improves the robustness of the system in complex scenarios such as occlusion and dynamic obstacles.

[0055] 2. This invention employs an improved RRT* algorithm and a dynamic potential field collaborative optimization mechanism, combined with GPU parallel acceleration technology, to achieve millisecond-level collision-free path replanning. At the same time, it uses point cloud normal vector matching technology to constrain the spray trajectory and the geometric orientation of the crop surface, avoiding drift loss of pesticide due to non-perpendicular incident, thus improving the targeting accuracy of pesticide application while ensuring obstacle avoidance safety.

[0056] 3. This invention constructs a droplet motion model based on unsteady fluid dynamics equations and electrostatic field control equations, and dynamically adjusts the high-voltage pulse voltage and electrode activation strategy in conjunction with real-time environmental parameters, so that charged droplets can still be directionally adsorbed onto the target surface under airflow disturbance, solving the problem of uneven deposition caused by environmental interference in traditional electrostatic spraying and significantly reducing pesticide residues.

[0057] 4. This invention uses the real-time inversion of the drug deposition distribution by the change rate of lidar reflection intensity, and combines it with a multi-objective reinforcement learning model to optimize path parameters and spray control commands online, forming a closed-loop control architecture of "perception-decision-execution-feedback", which breaks through the bottleneck of drug waste and insufficient coverage caused by the fixed parameters in traditional open-loop operations.

[0058] 5. This invention integrates four-dimensional environment modeling, dynamic obstacle avoidance, and spray control into an embedded hardware platform based on GPU-accelerated tensor decomposition, path planning, and fluid simulation algorithms. This meets the real-time and low-power requirements of farmland operation scenarios and provides a feasible solution for the intelligent upgrading of large-scale agricultural machinery. Attached Figure Description

[0059] Figure 1 This is a flowchart of the method of the present invention;

[0060] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0061] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Please see Figure 1 This invention provides an agricultural spraying optimization method based on lidar obstacle avoidance, comprising the following steps:

[0063] S1. Acquire three-dimensional point cloud data of farmland through lidar, and simultaneously collect thermal imaging data, attitude data and meteorological data to obtain multimodal environmental data;

[0064] In this embodiment, step S1 achieves full-dimensional perception of the farmland environment through multi-sensor collaborative acquisition and data fusion technology. The specific implementation process is as follows:

[0065] A lidar system is used to perform multi-beam scanning of a farmland area, generating a 3D point cloud dataset containing spatial coordinates and timestamps. The lidar emits laser pulses at a preset scanning frequency, and after receiving and analyzing the reflected signals, each point is obtained. spatial coordinates and corresponding timestamp This forms the initial three-dimensional point cloud set. ,in, .

[0066] Furthermore, the original point cloud data is subjected to noise reduction and sparsification processing. The point cloud is divided into uniform spatial units by a voxel grid filtering algorithm, retaining the point with the highest reflection intensity in each unit, eliminating redundant data and improving computational efficiency.

[0067] For each point in the point cloud data Searching for its radius neighborhood based on the KD-Tree data structure A local neighborhood set is formed by a number of neighboring points. Calculate the centroid coordinates of the neighborhood point set. Construct the covariance matrix :

[0068] ;

[0069] For the covariance matrix Eigenvalue decomposition yields three eigenvalues. and the corresponding feature vectors The smallest eigenvalue corresponding feature vector As a point normal vector It is used to characterize the local geometric orientation of crop surfaces.

[0070] While the lidar is scanning, environmental parameters are simultaneously collected using the following sensors:

[0071] Thermal imaging data: The thermal imager captures the infrared radiation intensity of farmland areas at a fixed frame rate, generating a temperature distribution matrix. Each element represents the temperature value at the corresponding coordinate point. Temperature data is aligned with point cloud spatial coordinates through coordinate system calibration, enabling cross-modal data mapping.

[0072] Attitude data: The inertial measurement unit (IMU) measures the pitch angle of the spray equipment in real time. Roll angle and yaw angle This is used to compensate for point cloud distortion caused by equipment movement. Preferably, a Kalman filter is used to dedop the original attitude angle data, improving data stability.

[0073] Meteorological data: Ambient wind speed was collected by an ultrasonic anemometer and a temperature and humidity sensor, respectively. and relative humidity By using a timestamp synchronization mechanism, meteorological parameters are associated with point cloud data to form a time-consistent multimodal environmental dataset.

[0074] The above 3D point cloud data and normal vector set Temperature distribution matrix Attitude angle and meteorological parameters The data is integrated according to a unified time base to construct a multi-dimensional environment data package. This package is then transmitted to the data processing module via a high-speed bus, providing input for subsequent environment modeling and path planning.

[0075] S2. Based on the multimodal environmental data, construct a four-dimensional environmental tensor that integrates spatial coordinates, time series, laser reflection intensity, and thermal imaging temperature. Reduce the dimensionality of the four-dimensional environmental tensor through tensor decomposition and fill in the missing regional data caused by occlusion to generate a passable region model.

[0076] In this embodiment, step S2 achieves a spatiotemporal consistent representation of the farmland environment through high-dimensional tensor modeling and missing data completion techniques. The specific implementation process is as follows:

[0077] Based on the multimodal environmental data obtained in step S1, four types of heterogeneous data—spatial coordinates, time series, laser reflection intensity, and thermal imaging temperature—are spatiotemporally aligned and structurally integrated. The spatial coordinate dimension is mapped from the 3D point cloud to uniformly divided grid cells through rasterization processing, forming a spatial dimension index. The time series dimension is divided according to a preset time window (e.g., 1 second) to generate a timestamp index. Laser reflection intensity and thermal imaging temperature are quantized into discrete levels, forming a reflection intensity dimension. With temperature dimension Finally, a four-dimensional environment tensor is constructed. Preferably, the quantization of reflection intensity and temperature adopts a non-uniform hierarchical strategy to adapt to the non-linear distribution characteristics of light and temperature in agricultural scenarios.

[0078] For the four-dimensional environment tensor Perform Tucker decomposition and represent it as a core tensor. With factor matrix The product form:

[0079] ;

[0080] Among them, the core tensor The dimension is determined according to the energy conservation criterion, satisfying Factor matrix , , , These correspond to low-rank projections for spatial, temporal, reflection intensity, and temperature dimensions, respectively. The decomposition parameters are iteratively optimized using alternating least squares (ALS) to achieve data dimensionality reduction and noise suppression.

[0081] Define the missing region index set ,in, Indexed by spatial coordinate dimensions; Indexed for time series dimensions; Indexed by the dimension of reflection intensity; Indexing the temperature dimension; constructing a low-rank tensor completion optimization problem:

[0082] ;

[0083] Among them, tensor nuclear norm The regularization coefficient represents the sum of the nuclear norms of the modal expansion matrices. To balance reconstruction error and low-rank constraints, the Alternating Direction Multiplier Method (ADMM) is used for iterative solution. By introducing auxiliary variables and Lagrange multipliers, the optimization problem is decomposed into the following sub-steps:

[0084] Tensor reconstruction subproblem: Fix other variables and update and complete the tensor. ;

[0085] Low-rank approximation subproblem: for the expansion matrix Perform a singular value thresholding (SVT) operation;

[0086] Multiplier update: Adjust the Lagrange multipliers according to the residuals.

[0087] The iteration terminates when the relative rate of change of the reconstruction error is less than a preset threshold, and the completed environment tensor is output. .

[0088] Based on the completed tensor The system extracts joint features from the spatial and temporal dimensions, and combines these features with reflection intensity and temperature thresholds (e.g., reflection intensity below the threshold indicates an obstacle, and areas with abnormal temperatures are marked as impassable) to generate a dynamically updated raster map of passable areas. Preferably, morphological dilation and erosion operations are used to smooth the raster map, eliminating the impact of isolated noise points on path planning.

[0089] S3. Based on the accessible area model, a collision-free path is planned using a dynamic potential field and path search algorithm. The collision-free path includes a spray trajectory that matches the crop surface normal vector.

[0090] In this embodiment, step S3 achieves collision-free path planning and directional spray trajectory control through the coordinated optimization of dynamic potential field and path search algorithm, combined with crop surface geometric feature matching technology. The specific implementation process is as follows:

[0091] Based on a traversable region model, a composite potential field function incorporating both gravitational and repulsive fields is constructed. The gravitational field guides the spraying device towards the target point, and its dynamic potential field function is defined as:

[0092] ;

[0093] Among them, the gravitational field , The current coordinates; The coordinates of the target point; The coordinates of the obstacle; The distance is Euclidean. This is the gain coefficient; This represents the maximum distance at which the obstacle can have an effect.

[0094] The repulsive field Used to avoid obstacles, its potential energy function is defined in piecewise form:

[0095] ;

[0096] Preferably, the repulsive field employs an adaptive gain strategy, adjusting according to the dynamic movement speed of the obstacle. The value is used to balance path smoothness and obstacle avoidance response speed.

[0097] Deploy the improved RRT* algorithm within a GPU parallel computing framework, generating an initial set of path nodes through parallel threads. With edge set During the node expansion phase, a potential field gradient direction biasing mechanism is introduced:

[0098] Random sampling point generation: Randomly sample candidate points within the passable area. ;

[0099] Potential field gradient guidance: Calculate the current potential field gradient direction ,right Apply a bias along the gradient descent direction to generate new nodes. ;

[0100] Collision detection and node selection: Calling the collision detection function Based on the completed environment tensor verify Whether it is located in a passable area, if the detection passes, then Add to the node collection.

[0101] When a dynamic obstacle position update is detected, local path replanning is triggered synchronously according to the multimodal environment data acquisition cycle. Preferably, an incremental tree structure adjustment strategy is adopted, optimizing only the affected path segments to reduce computational overhead. The replanning process inherits historical path node information and dynamically updates the repulsion field parameters. and This ensures the real-time performance and security of the path.

[0102] Normal vectors extracted from 3D point cloud data Calculate the spray direction vector With normal vector spatial angle :

[0103] ;

[0104] By jointly adjusting the pitch and yaw angles of the multi-degree-of-freedom robotic arm, it is possible to... (Preferred) (Set to 15°), where, A preset angle threshold is used to ensure that the pesticide droplets adhere vertically to the crop surface. The adjustment process employs a closed-loop PID control strategy to compensate in real time for spray direction deviations caused by equipment vibration or wind speed disturbances.

[0105] S4. Based on the collision-free path and real-time environmental parameters, combined with the drug diffusion model constructed based on the unsteady fluid dynamics equation and the electrostatic adsorption control parameters dynamically adjusted based on the three-dimensional point cloud data, three-dimensional directional spraying is performed. The electrostatic adsorption control parameters include a pulsed high voltage dynamically adjusted according to the target distance and wind speed.

[0106] In this embodiment, step S4 achieves precise control of three-dimensional directional spraying of the liquid medicine by coupling a fluid dynamics and electrostatic field control model and combining it with real-time environmental perception data. The specific implementation process is as follows:

[0107] Based on the coupling principle between the Navier-Stokes equations and the electrostatic field, a set of governing equations for the motion of drug droplets is established. This set of governing equations includes the mass conservation equation, the momentum conservation equation, and the electrostatic field equation, specifically in the following form:

[0108] ;

[0109] in, The density of the liquid medicine; For the velocity field; For pressure; Dynamic viscosity; The electrostatic field strength; It is the dielectric constant; Let be the charge density. The equations are solved using spatial discretization via the finite volume method, and the time progression employs an implicit Euler scheme to ensure numerical stability.

[0110] The wind speed collected in step S1 With humidity As external input parameters, the boundary conditions of the fluid model are dynamically modified:

[0111] Wind speed influence: Introducing a wind speed convection term into the momentum equation This characterizes the disturbance of the ambient airflow on the diffusion of the drug solution;

[0112] Humidity correction: Adjusted using a humidity-dependent charge decay model. ,Right now ,in, This is the humidity attenuation coefficient.

[0113] Based on target distance With real-time wind speed Calculate the amplitude of the pulsed high voltage. :

[0114] ;

[0115] in, The environmental correction factor was determined through calibration experiments. The target crop surface charge is related to the crop species and leaf surface characteristics. The high voltage is generated by a pulse generator, and its frequency is adaptively adjusted according to the droplet size distribution. Preferably, pulse width modulation (PWM) technology is used to achieve independent control of voltage amplitude and frequency.

[0116] Based on the point cloud normal vectors extracted in step S1 Calculate the direction vector of each ring electrode. Angle with the normal vector :

[0117] ;

[0118] when (Preferred) When set to 30°, the corresponding electrode is activated to generate a directional electrostatic field, causing charged droplets to adhere to the crop surface along the normal direction. The activation state of the electrode array is switched in real time via a high-speed switching circuit to ensure that the direction of the electrostatic field is adjusted synchronously with the spray trajectory.

[0119] The spray actuator receives the trajectory coordinates and nozzle attitude angle output by the path planning module, driving the multi-degree-of-freedom robotic arm and the fan-shaped nozzle to move in coordination. The liquid flow rate and atomization pressure are determined based on the real-time concentration distribution. Dynamic adjustment, while the electrostatic adsorption parameters are linked with the fluid model output through the feedback controller, forming a closed-loop control of "fluid diffusion-electrostatic adsorption-mechanical motion".

[0120] S5. Calculate the amount of drug deposition in real time by measuring the change rate of lidar reflection intensity, obtain the deposition results, and dynamically adjust the path planning and spray parameters based on the deposition results;

[0121] In this embodiment, step S5 utilizes lidar reflection characteristic analysis and multi-objective optimization control technology to achieve real-time monitoring and closed-loop parameter adjustment of drug deposition. The specific implementation process is as follows:

[0122] Acquire lidar reflection intensity data before and after spraying operations. and The rate of change of reflection intensity was calculated by comparing each point:

[0123] ;

[0124] The rate of change of reflection intensity Characterizing changes in surface optical properties caused by drug adhesion. An offline calibration experiment was established. With the amount of drug solution deposited Linear mapping relationship:

[0125] ;

[0126] Among them, calibration coefficient The calibration process is jointly determined by the optical absorptivity of the pesticide solution, the droplet size distribution, and the reflectivity of the crop surface, and is established through a reflectivity-deposition amount control experiment under controlled laboratory conditions. Preferably, the calibration process employs multispectral lidar to enhance adaptability to different pesticide solution types.

[0127] Real-time calculation of drug deposition volume Input a multi-objective reinforcement learning model, and define the state space as the current path planning parameters. Spray parameters and environmental parameter set The motion space is the parameter adjustment amount. , The optimization objectives of the model include:

[0128] Minimize waste of medicine solution: ,in, This represents the cumulative spray volume.

[0129] Maximize coverage: ,in, The area of ​​the region where the deposition exceeds the threshold; The total area of ​​the target work area.

[0130] Design reward function To balance multiple conflicting objectives:

[0131] ;

[0132] Among them, the weighting coefficient and Satisfy normalization constraints The policy optimization direction is dynamically adjusted by adjusting the preference weights. Preferably, the policy network is trained using the proximal policy optimization (PPO) algorithm, and the convergence speed of the model is improved by combining offline pre-training with online fine-tuning.

[0133] The reinforcement learning model outputs adjusted path planning parameters. (Including path curvature, travel speed) and spray parameters (Including flow rate, atomization pressure, and electrostatic voltage). Parameter adjustment commands are sent to the path planning module (step S3) and the spray actuator (step S4) via the real-time control bus, forming a closed-loop control circuit of "perception-decision-execution-feedback". Preferably, a sliding time window mechanism is used to smooth historical deposition data, eliminating the interference of instantaneous noise on parameter adjustment.

[0134] Please see Figure 2 The present invention also provides an agricultural spraying optimization system based on lidar obstacle avoidance, the system comprising:

[0135] The multimodal data acquisition module integrates a lidar, thermal imager, inertial measurement unit (IMU), and meteorological sensors. Through spatiotemporal synchronous acquisition by multiple heterogeneous sensors, it achieves real-time acquisition of three-dimensional point cloud data of the farmland environment, temperature field distribution, equipment motion attitude, and meteorological parameters (wind speed, humidity). The lidar uses a multi-beam scanning mode to generate dense point clouds; the thermal imager inverts crop surface temperature through infrared radiation intensity; the IMU provides three-axis acceleration and angular velocity data of the equipment to compensate for motion distortion; and the meteorological sensors communicate with the main control unit in real time via a bus protocol. Data from all sensors are aligned using a unified timestamp to form a multimodal environmental perception dataset.

[0136] The four-dimensional environment modeling module receives input from the multimodal data acquisition module and integrates spatial coordinates, time series, laser reflection intensity, and thermal imaging temperature into a four-dimensional environment tensor. It uses tensor decomposition technology to reduce the dimensionality of the high-dimensional data, extracts spatiotemporal correlation features, and employs a low-rank completion algorithm to repair missing data caused by vegetation occlusion or sensor blind spots. The module outputs a rasterized walkable area model, annotating obstacle boundaries, crop distribution, and temperature anomaly areas, providing high-confidence environmental prior information for path planning.

[0137] The dynamic path planning module, based on an improved RRT* algorithm and a GPU parallel acceleration architecture, plans collision-free paths in real time within a passable area model. The algorithm introduces a dynamic potential field guidance mechanism, generating path node bias directions through the superposition of gravitational and repulsive fields, and rapidly expanding the path tree using GPU multi-threaded parallel computation. The path planning results include the motion trajectory of the spraying equipment and a sequence of nozzle attitude angles matching the crop surface normal vector, ensuring that the spray direction is perpendicular to the target surface.

[0138] The directional spray control module integrates a non-steady-state hydrodynamic solver, a ring electrode array, and a high-voltage pulse generator to achieve coordinated control of pesticide diffusion and electrostatic adsorption. The hydrodynamic solver simulates the three-dimensional diffusion trajectory of droplets based on real-time environmental parameters (wind speed, humidity), the ring electrode array activates a directional electrostatic field based on the crop surface normal vector, and the high-voltage pulse generator dynamically adjusts the voltage amplitude and frequency to optimize droplet adsorption efficiency. All components achieve microsecond-level response via a control bus, ensuring strict synchronization between spray parameters and path planning.

[0139] The closed-loop feedback module calculates the amount of pesticide deposited in real time through a lidar reflection intensity analysis unit, and dynamically optimizes path planning parameters (such as travel speed and path curvature) and spray parameters (such as flow rate and voltage) by combining a reinforcement learning decision-making unit. The reinforcement learning model uses deposition uniformity and pesticide utilization rate as multi-objective reward functions, and generates parameter adjustment instructions through online strategy iteration, forming a closed-loop control loop of "perception-decision-execution".

[0140] The actuator module employs a multi-degree-of-freedom robotic arm and a high-precision fan-shaped nozzle. It performs directional spraying operations based on the trajectory coordinates output by the dynamic path planning module and the adjustment parameters of the directional spraying control module. The robotic arm analyzes the target pose using inverse kinematics algorithms, combined with a servo motor to achieve millimeter-level positioning accuracy. The fan-shaped nozzle uses piezoelectric ceramics to drive and adjust the atomization angle and flow rate, adapting to the spraying needs of different crop canopies. Each actuator communicates with the main control system via real-time Ethernet to ensure precise synchronization of operational commands.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An agricultural spraying optimization method based on lidar obstacle avoidance, characterized in that, Includes the following steps: S1. Acquire three-dimensional point cloud data of farmland through lidar, and simultaneously collect thermal imaging data, attitude data and meteorological data to obtain multimodal environmental data; S2. Based on the multimodal environmental data, construct a four-dimensional environmental tensor that integrates spatial coordinates, time series, laser reflection intensity, and thermal imaging temperature. Reduce the dimensionality of the four-dimensional environmental tensor through tensor decomposition and fill in the missing regional data caused by occlusion to generate a passable region model. S3. Based on the accessible area model, a collision-free path is planned using a dynamic potential field and path search algorithm. The collision-free path includes a spray trajectory that matches the normal vector of the crop surface. S4. Based on the collision-free path and real-time environmental parameters, combined with the drug diffusion model constructed based on the unsteady fluid dynamics equation and the electrostatic adsorption control parameters dynamically adjusted based on the three-dimensional point cloud data, three-dimensional directional spraying is performed. The electrostatic adsorption control parameters include a pulsed high voltage dynamically adjusted according to the target distance and wind speed. S5. Calculate the amount of drug deposition in real time by using the change rate of lidar reflection intensity to obtain the deposition results, and dynamically adjust the path planning and spray parameters based on the deposition results.

2. The agricultural spraying optimization method based on lidar obstacle avoidance according to claim 1, characterized in that, Step S1 includes the following steps: A 3D point cloud dataset is generated by scanning the farmland environment with LiDAR. , where each point Includes spatial coordinates and timestamp ; For each point in the three-dimensional point cloud data Select its local neighborhood point set Calculate the covariance matrix ,in, The neighborhood centroid; with respect to the covariance moments Perform eigenvalue decomposition, and take the eigenvector corresponding to the smallest eigenvalue as a point. normal vector ; Simultaneous acquisition of temperature distribution data output by thermal imager IMU-measured device attitude angle Wind speed collected by meteorological sensors and humidity This data is integrated into multimodal environmental data.

3. The agricultural spraying optimization method based on lidar obstacle avoidance according to claim 1, characterized in that, Step S2 includes the following steps: The spatial coordinates in the three-dimensional point cloud data timestamp Reflection intensity and thermal imaging temperature Integrate into a four-dimensional environment tensor according to a preset time window. ,in, For spatial coordinate dimensions; For time series dimensions; For the dimension of reflection intensity, Temperature dimension; The four-dimensional environment tensor is subjected to Tucker decomposition to obtain the core tensor. sum factor matrix ,satisfy .

4. The agricultural spraying optimization method based on lidar obstacle avoidance according to claim 3, characterized in that, In step S2, the process of completing the missing data due to occlusion includes: Based on the core tensor sum factor matrix Define the missing region index set ,in, Indexed by spatial coordinate dimensions; Indexed for time series dimensions; Indexed by the dimension of reflection intensity; Indexing the temperature dimension; constructing an optimization problem: ； in, For the tensor nuclear norm, The regularization coefficient is used; the optimization problem is solved iteratively using the alternating direction multiplier method, and the completed environment tensor is output. .

5. The agricultural spraying optimization method based on lidar obstacle avoidance according to claim 1, characterized in that, Step S3, based on the passable area model, uses a dynamic potential field and path search algorithm to plan a collision-free path, including: Constructing the dynamic potential function Among them, the gravitational field repulsive field satisfy: ； in, The current coordinates; The coordinates of the target point; The coordinates of the obstacle; The distance is Euclidean. This is the gain coefficient; This represents the maximum distance at which the obstacle can have an effect. Based on the improved RRT* algorithm, an initial path node set is generated within a GPU parallel computing framework. and edge set ; During the node expansion phase, based on the potential field gradient direction... Bias the random sampling points to generate new nodes. ; through collision detection function Verify node security, among which, The completed environment tensor; when a dynamic obstacle position update is detected, local path replanning is triggered with a response period synchronized with the multimodal environment data acquisition.

6. The agricultural spraying optimization method based on lidar obstacle avoidance according to claim 1, characterized in that, Step S3, where the collision-free path includes a spray trajectory that matches the crop surface normal vector, is achieved through the following steps: The normal vector extracted from the 3D point cloud data Calculate the spray direction vector With normal vector The included angle According to the included angle Adjust the pitch and yaw angles of the spray nozzles to make... ,in, This is a preset angle threshold.

7. The agricultural spraying optimization method based on lidar obstacle avoidance according to claim 1, characterized in that, In step S4, the implementation of the drug diffusion model based on the unsteady hydrodynamic equations includes: Establish the governing equations that couple the Navier-Stokes equations with the electrostatic field: ； in, The density of the liquid medicine; For the velocity field; For pressure; Dynamic viscosity; The electrostatic field strength; It is the dielectric constant; The charge density is used; the governing equations are discretized using the finite volume method, combined with real-time environmental parameters such as wind speed. With humidity Iteratively solve the three-dimensional drug concentration distribution .

8. The agricultural spraying optimization method based on lidar obstacle avoidance according to claim 1, characterized in that, In step S4, the dynamic adjustment of the electrostatic adsorption control parameters includes: Based on target distance With real-time wind speed Calculate the pulsed high voltage amplitude using the following formula. : ； in, This is an environmental correction factor; The surface charge of the target crop; the normal vector extracted from the three-dimensional point cloud data. Calculate the direction vector of each ring electrode. With normal vector The included angle ,when The corresponding electrode is activated to generate a directional electrostatic field.

9. The agricultural spraying optimization method based on lidar obstacle avoidance according to claim 1, characterized in that, Step S5 includes the following steps: Acquire lidar reflection intensity data before and after spraying. and Calculate the rate of change of reflection intensity ; The rate of change of reflection intensity was established through calibration experiments. With the amount of drug solution deposited mapping relationship ,in, These are calibration coefficients; The amount of drug solution deposited is input into a multi-objective reinforcement learning model to minimize the drug solution waste rate. and maximize coverage To optimize the objective, output the adjusted path planning parameters. and spray parameters The reward function of the reinforcement learning model is: ； in, These are the weighting coefficients.

10. An agricultural spray optimization system based on lidar obstacle avoidance, applied to the method described in any one of claims 1-9, characterized in that, The system includes: The multimodal data acquisition module integrates lidar, thermal imager, IMU and meteorological sensors to simultaneously acquire 3D point cloud data, temperature distribution, equipment attitude and environmental parameters; The four-dimensional environment modeling module is communicatively connected to the multimodal data acquisition module and is used to construct and complete the four-dimensional environment tensor to generate a passable area model. The dynamic path planning module, based on the improved RRT* algorithm and GPU parallel computing unit, is used to plan collision-free paths in real time. The directional spray control module includes an unsteady hydrodynamic solver, a ring electrode array, and a high-voltage pulse generator, which are used to perform three-dimensional drug diffusion simulation and dynamic control of electrostatic adsorption parameters. The closed-loop feedback module, through the lidar reflection intensity analysis unit and the reinforcement learning decision unit, is used to calculate the amount of drug deposition and optimize the path and spray parameters; The actuator module, which includes a multi-degree-of-freedom robotic arm and a fan-shaped nozzle, is used to perform directional spraying operations according to adjusted parameters.

Images