A Prediction System and Method for the Microenvironment of Flying Seeds Based on Sand Grain Motion Simulation

Abstract

This invention relates to a micro-environment prediction system and method for aerial seeding based on sand grain motion simulation, belonging to the field of UAV aerial seeding technology and environmental simulation prediction. The invention proposes to collect environmental data through a multi-source heterogeneous microscale sensing network, predict flow field data using a graph neural network-driven sand grain-airflow coupling simulation engine, and then calculate seed trajectory and displacement resistance coefficient through a seed-sand multi-physics field coupling behavior prediction model, generating a displacement risk probability distribution map. An adaptive aerial seeding parameter intelligent optimization system optimizes hovering height and ejection parameters accordingly, and achieves precise ejection through a variable stiffness ejection actuator. The system significantly improves the seed displacement resistance coefficient by 2.8 times, effectively addressing the problem of seed displacement in windy and sandy environments, improving the accuracy and success rate of UAV aerial seeding, and providing efficient technical support for desertification control.

Description

Technical Field

[0001] This invention relates to a microenvironment prediction system and method for aerial seeding based on sand grain movement simulation, belonging to the field of UAV aerial seeding technology and environmental simulation prediction. Background Technology

[0002] Aerial seeding is an important means of restoring vegetation and preventing desertification. Traditional aerial seeding technologies, such as the dual-mode intelligent aerial seeding system disclosed in CN 118270236A for precision aerial seeding operations, include an intelligent drone, a multi-agent control platform, a seedling-bearing aerial seeding terminal, a cold-launch system, and a hole-digging robot, as well as the collaborative operation mechanism among them. This system improves the overall efficiency of aerial seeding operations through high-precision navigation and positioning, dual-mode operation, and intelligent control.

[0003] However, existing technologies have the following obvious shortcomings: First, they lack the ability to predict the microenvironment, making it impossible to accurately predict the microenvironmental conditions after the seeds land, especially since seeds are easily blown away and displaced in windy and sandy environments; second, the simulation of seed landing trajectories is insufficient, lacking the ability to simulate the movement trajectory of seeds under the action of wind in real time; third, the hovering height is fixed, making it impossible to adaptively adjust the optimal sowing height according to the real-time wind field environment; fourth, the seed stability is poor, as traditional systems mainly focus on the accuracy of seed placement and neglect the stability guarantee after the seeds land; and fifth, there is the problem of computational delay, lacking the ability to respond quickly in complex wind field environments.

[0004] The aforementioned shortcomings result in low efficiency of aerial seeding in windy and sandy environments and low seed survival rates, posing a significant challenge, especially for aerial seeding afforestation in arid and semi-arid regions. Summary of the Invention

[0005] The purpose of this invention is to provide a microenvironment prediction system and method for aerial seeding based on sand particle motion simulation. Through multi-scale microenvironment perception and intelligent prediction, it solves the technical problem of seed displacement in windy and sandy environments, and achieves precise seed delivery and stable landing.

[0006] This invention proposes a microenvironment prediction system for aerial seeding based on sand grain movement simulation, comprising:

[0007] The data acquisition module is used to acquire 3D point cloud data through multi-sensor fusion.

[0008] The preprocessing module is used to perform noise reduction and downsampling processing on the three-dimensional point cloud data;

[0009] The feature extraction module is used to extract features from the preprocessed point cloud data and generate semantic labels;

[0010] The reference plane acquisition module is used to calculate the probability optimal fitting plane based on the semantic labels;

[0011] The pixel feature conversion module is used to map the point cloud features to the pixel feature space to achieve multi-scale feature output.

[0012] Preferably, the data acquisition module includes a depth camera and a lidar, wherein the depth camera is a binocular camera and includes a data acquisition unit, an image stitching unit, an image correction unit, and a data storage unit. The depth camera and lidar are mounted on the drone for scanning and imaging the target work area.

[0013] Preferably, the preprocessing module includes:

[0014] The geometric denoising submodule is used to remove flying points generated during data acquisition. It uses a pass-through filtering algorithm and a voxel grid filtering algorithm for data downsampling.

[0015] The outlier removal submodule is used to filter out outliers outside the point cloud distribution area using a statistical filtering algorithm, reducing the number of subsequent calculations.

[0016] Preferably, the feature extraction module is implemented using the PointNet++ convolutional computing network, which includes point cloud feature extraction, feature upsampling, feature fusion and semantic label generation processes. The module uses a local feature enhancement mechanism and a context-aware module to capture the hierarchical structure information of the point cloud.

[0017] Preferably, the feature extraction module further includes an adversarial generative network (PGN) for receiving the semantic label vector and generating lithology labels, wherein the GPN includes:

[0018] A generator is used to convert semantic label vectors into lithological distribution features;

[0019] The discriminator is used to distinguish between generated labels and real labels and continuously optimizes the generated features through error feedback.

[0020] Preferably, the reference surface acquisition module is used for:

[0021] Calculate the proportion of point cloud numbers for each semantic category to the total number of point clouds;

[0022] When the ratio is not greater than a preset threshold, feature extraction is performed on each semantic category, including the centroid, normal vector, local surface normal vector, covariance matrix, point pairs, principal components, and distance from the point to the fitted surface.

[0023] The semantic categories and features were fitted using the hybrid RANSAC algorithm to obtain multiple sets of different semantic planes and the confidence probabilities of each semantic plane;

[0024] The final reference surface is determined based on the confidence probability.

[0025] Preferably, the pixel feature conversion module is implemented using the Unet architecture, which includes two parts: encoding and decoding. The encoding part consists of convolution, pooling, and downsampling, while the decoding part is implemented using transposed convolution. Skip connections are used in the middle of the Unet architecture to fuse the high-level semantics at each scale during the encoding process with the detailed features at the low level of the decoding process.

[0026] Preferably, the adversarial generative network learns based on physical constraints, including:

[0027] Constructing adversarial learning between semantic and physical models;

[0028] Design a new semantic segmentation model label based on terrain label information and color map information;

[0029] The semantic model is optimized using the cross-entropy loss function to enable it to learn physical properties.

[0030] A progressive constraint strategy is adopted, focusing on feature learning in the early stage of training and gradually increasing the weight of physical constraints in the later stage.

[0031] Preferably, the formula for calculating the plane confidence level by the reference plane acquisition module is:

[0032] Confidence level = semantic consistency × geometric fit × physical plausibility

[0033] Among them, semantic consistency represents the degree of consistency of semantic labels of points in the plane, geometric fit represents the residual error of plane fitting, and physical rationality represents the rationality of evaluating the plane based on physical rules.

[0034] The method based on the aforementioned 3D point cloud semantic segmentation and intelligent fitting system for measurement reference surfaces includes:

[0035] The steps for acquiring 3D point cloud data involve using a depth camera and LiDAR to scan and image the target work area to obtain 3D point cloud data.

[0036] The preprocessing step involves denoising and downsampling the 3D point cloud data;

[0037] In the feature extraction step, the PointNet++ network is used to extract features from the preprocessed point cloud data and generate semantic label vectors.

[0038] The lithological label generation step involves inputting the semantic label vector into a generative adversarial network to generate lithological labels.

[0039] The benchmark surface acquisition step involves calculating the probabilistic optimal fitting plane based on the lithological label, determining the confidence level of the plane, and if the confidence level is higher than a preset threshold, then the plane is determined as the measurement benchmark surface.

[0040] In the online inference step, the three-dimensional point cloud data is input into the trained model for processing to obtain the measurement reference surface within the target area.

[0041] The beneficial effects of this invention are as follows:

[0042] 1. Significantly improves the seed displacement resistance coefficient by 2.8 times, effectively solving the core problem of seed displacement in windy and sandy environments;

[0043] 2. The seed accuracy rate has been increased to 92%, a 15% improvement compared to traditional techniques;

[0044] 3. Seed survival rate increased to 85%, a 25% improvement compared to traditional techniques;

[0045] 4. The system's adaptability to wind speeds has been expanded to 0-12 m / s, significantly extending the window period for aerial seeding operations;

[0046] 5. Sowing efficiency increased by 30%, and the daily operating area of ​​a single drone increased to 150 mu. Attached Figure Description

[0047] Figure 1 This is a diagram showing the overall architecture of the microenvironment prediction system for aerial seeding based on sand particle motion simulation, as described in this invention.

[0048] Figure 2 This is a schematic diagram of the structure of the multi-source heterogeneous microscale sensing network of the present invention;

[0049] Figure 3 This is a flowchart of the workflow of the neural network-driven sand-airflow coupling simulation engine of the present invention.

[0050] Figure 4 This is a schematic diagram of the seed-sand multiphysics coupling behavior prediction model of the present invention;

[0051] Figure 5 This is a structural diagram of the adaptive aerial seeding parameter intelligent optimization system of the present invention;

[0052] Figure 6 This is a structural diagram of the variable stiffness catapult actuator of the present invention;

[0053] Figure 7 This is a schematic diagram of the physical information enhancement graph neural network in the system of the present invention;

[0054] Figure 8 This is a visualization of the sand particle motion field prediction in the system of the present invention;

[0055] Figure 9 This is a schematic diagram illustrating the factors affecting the seed displacement resistance coefficient in the system of this invention;

[0056] Figure 10 This is a flowchart of the method of the present invention. Detailed Implementation

[0057] Please refer to the attached document. Figure 1-10 The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0058] Example 1

[0059] like Figure 1 As shown, the aerial seeding microenvironment prediction system based on sand particle motion simulation of the present invention includes a multi-source heterogeneous microscale sensing network 1, a graph neural network-driven sand particle-airflow coupling simulation engine 2, a seed-sand multi-physics field coupling behavior prediction model 3, an adaptive aerial seeding parameter intelligent optimization system 4, and a variable stiffness catapult actuator 5. These five modules constitute a closed-loop system of perception-prediction-optimization-execution, working together to achieve precise seed delivery and stable landing.

[0060] A multi-source heterogeneous microscale sensing network 1 is responsible for collecting wind field data, sand particle movement data, and surface characteristic data, and generating a standard environmental state matrix E(t). A graph neural network-driven sand particle-airflow coupling simulation engine 2 receives the standard environmental state matrix E(t) and generates flow field prediction data F(t+Δt) through a physical information-enhanced graph neural network. A seed-sand multi-physics coupling behavior prediction model 3 calculates the seed trajectory T(s) and seed displacement resistance coefficient k(s,E) based on the flow field prediction data F(t+Δt), and generates a displacement risk probability distribution map P(x,y). An adaptive aerial seeding parameter intelligent optimization system 4 calculates the optimal hovering height h* and ejection parameters (v,θ) based on the displacement risk probability distribution map P(x,y), seed trajectory T(s), and seed displacement resistance coefficient k(s,E), and generates control commands C(t). The variable stiffness catapult actuator 5 receives the control command C(t), adjusts the catapult stiffness k* and seed wrapping parameters, executes seed catapult, and collects the actual trajectory data Ta(s) as feedback data to send to the adaptive aerial seeding parameter intelligent optimization system 4.

[0061] Data transfer between system modules utilizes 100Mbps high-speed transmission with latency controlled within 10ms to ensure real-time performance. The entire system's workflow is kept within 500ms, enabling rapid response and decision-making.

[0062] Example 2

[0063] like Figure 2 As shown, the multi-source heterogeneous microscale sensing network 1 includes a micro PIV particle image velocimetry system 11, a micro six-degree-of-freedom wind field sensor array 12, a surface characteristic fast detector 13, and a data fusion processing unit 14.

[0064] The miniature PIV (Particle Image Velocimetry) system 11 consists of a 532nm pulsed laser sheet light source 111 and a 4000fps high-speed camera 112, achieving a spatial resolution of 0.5mm. It is used to capture the motion trajectory and velocity field of sand particles at the 10-500μm scale. The miniature PIV system employs dual-pulse synchronization technology, obtaining the motion vector of the sand particles through correlation analysis of consecutive frame images. The system is set with a light sheet thickness of 1mm, a pulse interval of 50μs, and a preferred laser power of 300mW, sufficient to illuminate the tiny sand particles without affecting the measurement.

[0065] The miniature six-degree-of-freedom wind field sensor array 12 consists of eight miniature hot-wire anemometers 121 arranged in a cubic array. It has a sampling frequency of 200Hz, a measurement accuracy of ±0.05m / s, and is used to measure wind speeds over a 3m radius. 3 The three-dimensional wind field distribution in space. The hot-wire anemometer operates in a constant-temperature mode, with a 50°C difference between its operating temperature and the ambient temperature to improve sensitivity. The sensor array calculates the wind speed vector at any point in space using Bulman interpolation, achieving a spatial resolution of 10 cm.

[0066] The rapid surface characteristic detector 13 includes a miniature laser roughness meter 131 and a capacitive moisture content sensor 132, used to acquire the roughness, density, and moisture content parameters of the surface sand layer. The laser roughness meter uses the triangulation principle, with a measurement range of 0-10 mm and an accuracy of 0.01 mm. The moisture content sensor operates at a frequency of 100 MHz, with a measurement range of 0% to 20% and an accuracy of 0.5%.

[0067] The data fusion processing unit 14 employs the Kalman filter algorithm to perform spatiotemporal alignment and fusion of multi-sensor data, generating a standard environmental state matrix E(t). This matrix contains wind field vectors, sand grain movement vectors, and surface characteristic parameters, with dimensions m×n, where m represents the number of spatial discrete points and n represents the feature dimension. The Kalman filter algorithm can be expressed as:

[0068]

[0069] in: F is the state estimate calculated at time k. k Let B be the state transition matrix. k For the control matrix, u k K is the control vector. k For Kelman gain, z k H is the observation vector. k This represents the observation matrix. The Kalman filter algorithm achieves optimal estimation of the environmental state through two steps: prediction and update. The filter update frequency is 100Hz, ensuring the real-time performance and accuracy of the environmental state matrix.

[0070] Example 3

[0071] like Figure 3 As shown, the graph neural network-driven sand-airflow coupling simulation engine 2 includes a physical information-enhanced graph neural network module 21, a multi-scale hybrid discrete element model 22, a recurrent convolutional GNN accelerator 23, and a fast flow field prediction algorithm module 24.

[0072] The Physical Information Augmented Graph Neural Network module 21 uses sand grains as nodes and flow fields as edges to construct a physical information augmented graph network structure. This structure integrates the Navier-Stokes equation constraints from traditional CFD into the message passing mechanism of the graph network, enabling fast computation under physical constraints. The graph network updates the nodes via the node update function φ. v and edge update function φ e To achieve information transmission:

[0073]

[0074] in: The features of the l-th layer node, The features of the l-th layer node, Let N(i) be the edge feature connecting nodes i and j, and N(i) be the set of neighbors of the node. v and φ e These are the update functions for nodes and edges, respectively. The node update function is implemented using a multilayer perceptron with a hidden layer dimension of 128 and a Leaky ReLU activation function. The edge update function also uses a multilayer perceptron, but with a hidden layer dimension of 64.

[0075] A multi-scale hybrid discrete element model22 combines the discrete element method (CEM) with computational fluid dynamics (CFD) to construct a three-phase interactive mathematical model of sand particles, seeds, and airflow. This model uses a gridded CFD method to solve the flow field at the macroscopic scale and the discrete element method to simulate sand particle motion at the microscopic scale, coupling the two scales through momentum exchange. The contact force between sand particles in the model is calculated using the Hertz-Mindlin contact model, with a damping coefficient of 0.3, a static friction coefficient of 0.6, and a dynamic friction coefficient of 0.4.

[0076] The Recurrent Convolutional GNN accelerator23, powered by a pre-trained model, compresses traditional 1-2 minute CFD computations to within 50ms. The accelerator employs a 4-layer recurrent convolutional GNN network, with each layer containing 32 convolutional kernels, a stride of 1, and SAME padding. The model is pre-trained on 50,000 sets of CFD simulation data using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. Model compression utilizes Tucker decomposition technology, achieving a compression ratio of 85% while retaining 98% of the prediction accuracy.

[0077] The flow field fast prediction algorithm module 24 adopts dynamic adaptive mesh generation technology to achieve 3m 3Prediction of sand grain movement field within a spatial range with a spatial resolution of 2cm. 3 The time resolution reaches 50ms. The algorithm automatically refines the mesh in high-gradient regions and uses a coarse mesh in regions with gentle gradients, optimizing the allocation of computational resources. The mesh refinement threshold is set to 15% of the gradient value, the maximum refinement level is 3 layers, and the number of meshes is dynamically controlled to within 100,000.

[0078] Example 4

[0079] like Figure 4 As shown, the seed-sand multi-physics coupling behavior prediction model 3 includes a multi-physics coupling model module 31, a seed motion trajectory prediction algorithm module 32, a sand grain migration risk assessment engine 33, and a real-time displacement coefficient calculator 34.

[0080] The multiphysics coupling model module 31 integrates mechanical force fields, fluid force fields, and surface energy fields to construct a multiphysics coupling mathematical model. This model considers the gravity, buoyancy, fluid drag force, lift force, and the adhesion and friction forces between the seed and sand grains acting on the seed. The force balance equation can be expressed as:

[0081]

[0082] Where: m is the seed mass. The seed position vector, For gravity, For buoyancy, For drag force, For lift, For adhesion force, Frictional force is used. The drag force calculation employs a modified Drague formula, considering correction factors for Reynolds numbers ranging from 10 to 1000. The adhesion force model is based on JKR contact theory, taking into account the relationship between seed surface energy and sand grain contact area.

[0083] The seed trajectory prediction algorithm module 32 is based on the Lagrange particle tracking method to predict the trajectory of a seed from release to landing. The algorithm uses a fourth-order Runge-Kutta method with an adaptive time step to solve the motion equations, with an initial time step of 10 ms, a minimum time step of 0.1 ms, and a relative error threshold of 10⁻⁵. Considering the influence of seed geometry, the algorithm introduces a shape factor, and obtains the drag coefficient and lift coefficient for different seed shapes by fitting wind tunnel experimental data.

[0084] The sand grain migration risk assessment engine 33 establishes a risk assessment model based on Monte Carlo simulation to calculate the displacement probability distribution of seeds after they land under different wind speed conditions. The engine performs 10,000 random sampling simulations, considering random variations in factors such as wind speed, direction, surface roughness, and seed characteristics, to generate a displacement risk probability distribution map P(x,y). The risk levels are divided into three levels: low, medium, and high, corresponding to displacement probabilities of <20%, 20%–50%, and >50%, respectively.

[0085] The Real-Time Displacement Resistance Calculator 34 constructs a seed displacement resistance coefficient prediction model, comprehensively considering seed shape, mass, surface characteristics, and surface parameters. The displacement resistance coefficient k(s,E) is defined as an index of a seed's ability to resist wind-induced displacement, ranging from 0 to 1; a larger value indicates stronger displacement resistance. The calculation formula is:

[0086]

[0087] Where: m s For seed quality, A s F represents the windward surface area of ​​the seed. a For adhesion force, F d For wind-driven drag force, R s M represents the surface roughness of the seed. s The seed moisture content is represented by α, β, γ, and δ, which are weighting coefficients obtained by fitting experimental data and are 0.3, 0.4, 0.2, and 0.1, respectively.

[0088] Example 5

[0089] like Figure 5 As shown, the adaptive aerial seeding parameter intelligent optimization system 4 includes a hovering height dynamic optimizer 41, a multi-objective Bayesian optimization framework 42, a pressure-ejection relationship self-learning module 43, and a state space feedback controller 44.

[0090] The hovering height dynamic optimizer 41 utilizes a deep reinforcement learning algorithm to calculate the optimal hovering height within the range of 0.5–2.5 m in real time. The optimizer employs a deep Q-network (DQN) structure. The state space includes current wind field characteristics, surface properties, and seed parameters. The action space consists of discretized hovering height values ​​(0.5 m, 1.0 m, 1.5 m, 2.0 m, 2.5 m). The reward function comprehensively considers seed landing accuracy and anti-displacement coefficient. The network structure is a 3-layer fully connected network with 256, 128, and 64 hidden layer nodes, respectively. The activation function is ReLU, the learning rate is 0.001, the discount factor is 0.95, and the initial ε of the ε-greedy policy is 0.9 with a decay rate of 0.995.

[0091] A multi-objective Bayesian optimization framework was constructed to evaluate seed accuracy, displacement resistance, and efficiency, dynamically calculating the optimal seeding parameters. The framework uses a Gaussian process regression model to construct a surrogate model, employing Expected Improvement (EI) as the acquisition function. Optimization variables include launch velocity (1–10 m / s), launch angle (30°–60°), and seed wrapping parameters. The multi-objective optimization employed the Pareto front method, using the hypervolume index to evaluate the solution set quality. The kernel function for the Gaussian process was chosen to be the Matérn 5 / 2 kernel, with a length scale parameter of 0.1 and a signal variance of 1.0.

[0092] The air pressure-ejection relationship self-learning module 43 establishes the mapping relationship between air pressure and ejection velocity and angle. This module employs an online learning method, continuously updating model parameters based on actual ejection data. The mapping relationship uses a quadratic polynomial model:

[0093] v = a1P 2 +b1P+c1,

[0094] θ=a2P 2 +b2P+c2,

[0095] Where: v is the ejection velocity, θ is the ejection angle, P is the air pressure value, and a1, b1, c1, a2, b2, and c2 are model parameters. The initial parameters are obtained through pre-calibration. Online learning adopts the recursive least squares method, and the forgetting factor is set to 0.95 to ensure that the model can adapt to the dynamic changes in the air pressure-ejection relationship.

[0096] The state-space feedback controller 44 is based on a micro-environment state-space model to achieve real-time closed-loop control of parameters. The controller employs a linear quadratic regulator (LQR) design method. State variables include current position error, velocity error, and cumulative error, while the control variable is the air pressure regulation. The state-space equation is:

[0097]

[0098] y = Cx,

[0099] Where: x is the state vector, u is the control vector, y is the output vector, A is the system matrix, B is the input matrix, and C is the output matrix. The state feedback gain matrix K of the LQR controller is obtained by solving the Riccati equation. The state weight matrix Q is set as a diagonal matrix diag(10,5,1), the control weight matrix R is set to 1, and the control update frequency is 50Hz.

[0100] Example 6

[0101] like Figure 6As shown, the variable stiffness ejection actuator 5 includes a magnetorheological fluid-based variable stiffness ejection mechanism 51, a seed surface biomimetic microstructure processing unit 52, an environmentally responsive hydrogel encapsulation device 53, and an aerodynamic optimization processing unit 54.

[0102] The magnetorheological fluid-based variable stiffness ejection mechanism 51 dynamically adjusts the spring stiffness by regulating the magnetic field strength. The mechanism employs a flexible cavity structure filled with magnetorheological fluid. The magnetic field is generated by an electromagnetic coil with 500 turns and a maximum current of 2A, producing a magnetic field strength ranging from 0 to 1.2T. The magnetorheological fluid is prepared by suspending carbonyl iron powder (average particle size 5μm, volume fraction 30%) in silicone oil (viscosity 0.1Pa·s). The stiffness adjustment range is 5–50 N / m, the response time is less than 50 ms, and the relationship between stiffness and magnetic field strength is non-linear, achieving rapid mapping through a lookup table method. The compressed air storage tank has a volume of 500 mL and a maximum working pressure of 1.5 MPa. The air pressure is precisely controlled by a proportional pressure regulating valve, with a pressure adjustment range of 0.1–1.0 MPa and an accuracy of 0.01 MPa.

[0103] The seed surface biomimetic microstructure processing unit 52 constructs an array of wind-resistant microstructures on the seed surface. Based on the biomimetic design of the microstructures on the surface of desert plant seeds, the unit utilizes 3D micro / nano printing technology to fabricate microstructure templates with feature sizes ranging from 50 to 200 μm. The processing includes: cleaning the seed surface, coating with photocurable resin, imprinting the microstructures, UV curing, and post-treatment. The microstructure morphologies include conical arrays, grooved arrays, and hybrid structures, with densities ranging from 100 to 400 units / mm². 2 The height ranges from 30 to 120 μm. The microstructure increases the contact area between the seeds and sand grains, thus improving adhesion.

[0104] The environmentally responsive hydrogel coating device 53 provides seeds with a composite hydrogel coating layer containing silicates and alginates. The hydrogel formulation includes sodium alginate (2% w / v), sodium silicate (1% w / v), calcium chloride (1% w / v), and glycerol (5% w / v). The coating process employs fluidized bed technology, controlling the coating layer thickness within the range of 50-200 μm. The hydrogel exhibits strong water absorption and swelling properties, rapidly absorbing soil moisture upon landing to form a stable layer. Its water absorption rate can reach 20 times its own weight, with an expansion time of 10-30 seconds. Nutrients and moisturizers are also added to the coating layer to improve seed survival rates.

[0105] The aerodynamic optimization unit 54 optimizes the seed's aerodynamic shape, reducing the wind load ratio and improving stability. Based on previous wind tunnel experimental data, the unit constructs a seed aerodynamic characteristic database, containing drag and lift coefficients for 50 typical shapes. Optimization employs shape modification technology, adjusting the seed's center of gravity and aerodynamic characteristics by adding aerodynamic attachments. The attachments are made of biodegradable plant fibers, and their shapes include stabilizing fins, vortex generators, and drag rings, with their weight controlled to within 5% of the seed's weight. The optimization objective is to minimize the wind load ratio (wind force / gravity) while ensuring a stable descent attitude.

[0106] Example 7

[0107] In this embodiment, the key parameters of the system are set as follows:

[0108] The standard environmental state matrix E(t) has 1000×15 dimensions, including 1000 spatial discrete points and 15 feature dimensions (3D wind speed, 3D sand grain velocity, 3D location coordinates, 3D surface characteristic parameters, and 3D time stamp). The update frequency is 100Hz, and the data transmission rate is 100Mbps.

[0109] The spatial resolution of the flow field prediction data F(t+Δt) is 2 cm, the time prediction step is 50 ms, and the calculation time is controlled within 100 ms. The prediction accuracy is verified by comparison with high-precision CFD simulation, and the relative error is less than 5%.

[0110] The spatial coverage of the displacement risk probability distribution map P(x,y) is 10m×10m, with a resolution of 10cm×10cm, and the calculation is based on 10,000 Monte Carlo simulations. The risk level is divided into three levels: low risk (green, displacement probability <20%), medium risk (yellow, displacement probability 20%–50%), and high risk (red, displacement probability >50%).

[0111] The optimal hovering height h* has a search range of 0.5–2.5 m, a step size of 0.1 m, and an optimized response time of less than 150 ms. The ejection parameters include velocity v (range 1–10 m / s) and angle θ (range 30°–60°), with parameter resolutions of 0.1 m / s and 1°, respectively.

[0112] The control command C(t) includes hovering height, launch speed, launch angle, stiffness parameters, and wrapping parameters, encoded as a 32-byte command packet with a transmission delay of less than 5ms.

[0113] The system has a 500ms operating cycle, comprising four stages: sensing (100ms), prediction (100ms), optimization (100ms), and execution (200ms). The system can operate stably within a wind speed range of 0-12m / s, and is suitable for temperatures ranging from -20℃ to 55℃ and relative humidity from 10% to 90%.

[0114] Example 8

[0115] like Figure 7 As shown, the system also includes a closed-loop feedback and continuous optimization module 6, which is used to compare the deviation between the theoretical trajectory and the actual trajectory, calculate the correction factor, update the model parameters, and continuously optimize the decision-making strategy based on the reinforcement learning method.

[0116] The closed-loop feedback and continuous optimization module 6 includes a trajectory deviation analysis unit 61, a model adaptive adjustment unit 62, a reinforcement learning optimization unit 63, and a performance monitoring and evaluation unit 64.

[0117] The trajectory deviation analysis unit 61 compares the deviation between the theoretically predicted trajectory T(s) and the actual measured trajectory Ta(s) and calculates the correction factor c. The deviation is quantified using Euclidean distance and dynamic time warping (DTW) algorithms, and the correction factor is calculated using an exponentially weighted moving average method with a smoothing coefficient of 0.3. The deviation analysis considers three dimensions: position, velocity, and angle, generating a comprehensive deviation index.

[0118] The adaptive adjustment unit 62 updates the parameters of the prediction model based on the correction factor c. The adjustment uses gradient descent, with the learning rate dynamically adjusted according to the deviation, ranging from 0.001 to 0.1. Parameter updates are performed every 10 aerial seeding attempts to prevent overfitting. Key parameters adjusted include turbulence intensity parameters in the flow field prediction model, and adhesion and wind force coefficients in the seed-sand coupling model.

[0119] The reinforcement learning optimization unit 63 continuously optimizes the decision-making strategy based on actual results. The unit employs the Proximal Policy Optimization (PPO) algorithm, with both the policy and value networks using 3-layer fully connected networks. The hidden layer has 128 nodes, and the activation function is tanh. Training is conducted online, with each batch consisting of 500 fly-seeding experiences, and training lasts for 10 epochs. The reward function comprehensively considers seed landing accuracy, anti-displacement effect, and operational efficiency, with weights of 0.4, 0.4, and 0.2, respectively.

[0120] The performance monitoring and evaluation unit 64 monitors various system indicators in real time and generates a performance evaluation report. Monitoring indicators include seed landing accuracy (target point error), anti-displacement effect (displacement distance / wind speed), prediction accuracy (deviation between predicted and actual trajectories), and system response time. An evaluation report is generated every 100 aerial seeding attempts, including statistical analysis of the indicators and improvement suggestions.

[0121] The iteration cycle of the closed-loop feedback and continuous optimization module 6 is 500ms, and the optimization speed of a single iteration is no less than 15% per 100 fly-throughs. Through continuous learning and optimization, the system performance continuously improves during long-term use, adapting to different environmental conditions.

[0122] Example 9

[0123] This system also has a variety of extended functions to adapt to different application scenarios and environmental conditions.

[0124] First, the system supports parameterized configuration for different seed types. The built-in seed database contains physical characteristic parameters for 50 common afforestation and grass seed species, including geometric dimensions, mass, shape factor, surface roughness, and water absorption. Users can select seed types through the interface, and the system automatically loads the corresponding parameters and optimizes the aerial seeding strategy. For new seed types not included in the database, the system provides a rapid parameter measurement process, including image recognition measurement and physical characteristic testing, completing parameterization within 15 minutes.

[0125] Secondly, the system supports collaborative operations among multiple drones. A single control center can simultaneously manage 5-10 drones equipped with this system, assigning them to different work areas and monitoring their progress. Drones share environmental data and aerial seeding experience through a self-organizing network, improving overall operational efficiency. The collaborative strategy employs an auction-based task allocation mechanism, considering drone battery power, location, and payload to achieve balanced task load.

[0126] Furthermore, the system integrates a post-aerial seeding evaluation function. A high-resolution camera records the seed landing locations, and combined with GPS positioning information, a seeding distribution map is generated. Thirty days after seeding, the system can conduct a follow-up flight, using multispectral imaging technology to assess seed germination and growth, calculate survival rate and coverage, and provide a basis for optimizing subsequent aerial seeding strategies.

[0127] In addition, the system supports an emergency mode for extreme weather conditions. When the wind speed exceeds the threshold (12m / s) or a sudden strong wind occurs, the system automatically switches to a conservative mode, adjusting the aerial seeding parameters to ensure safe operation. In emergency mode, the system prioritizes seed resistance to displacement, appropriately sacrificing seeding accuracy by using a lower hovering height and stronger ejection force to ensure the seeds land quickly and securely.

[0128] Example 10

[0129] like Figure 10 As shown, the method for predicting the microenvironment of aerial seeding based on sand grain motion simulation includes the following steps:

[0130] Step S1: Collect wind field data, sand grain movement data, and surface characteristic data through a multi-source heterogeneous microscale sensing network, and generate a standard environmental state matrix E(t).

[0131] Specifically, a miniature PIV particle image velocimetry system is used to capture the trajectory and velocity field of sand particles, a miniature six-degree-of-freedom wind field sensor array is used to measure the three-dimensional wind field distribution, a surface characteristic rapid detector is used to obtain the physical parameters of the surface sand layer, and a Kalman filter algorithm is used to perform spatiotemporal alignment and fusion of multi-sensor data to generate a standard environmental state matrix.

[0132] Step S2: Input the standard environmental state matrix E(t) into the sand-airflow coupling simulation engine driven by the graph neural network, and generate flow field prediction data F(t+Δt) through the physical information-enhanced graph neural network.

[0133] Specifically, sand grains are used as nodes and the flow field as edges to construct a physical information-enhanced graph network structure. Combining the discrete element method and computational fluid dynamics, a three-phase interactive mathematical model of sand grains, seeds, and airflow is built. Fast inference is achieved through a pre-trained recurrent convolutional GNN network, and a dynamic adaptive mesh generation technique is employed to achieve a 3m... 3 Prediction of sand grain movement field within a spatial range.

[0134] Step S3: Based on the flow field prediction data F(t+Δt), calculate the seed motion trajectory T(s) and seed displacement resistance coefficient k(s,E) through the seed-sand multi-physics field coupled behavior prediction model, and generate the displacement risk probability distribution map P(x,y).

[0135] Specifically, a multi-field coupled mathematical model integrating mechanical force field, fluid force field and surface energy field is used to predict the trajectory of seeds from release to landing based on the Lagrange particle tracking method. A risk assessment model based on Monte Carlo simulation is established to calculate the displacement probability distribution of seeds under different wind speed conditions after landing. A seed displacement resistance coefficient prediction model is constructed, taking into account seed shape, mass, surface characteristics and surface parameters.

[0136] Step S4: Input the displacement risk probability distribution map P(x,y), seed motion trajectory T(s), and seed anti-displacement coefficient k(s,E) into the adaptive aerial seeding parameter intelligent optimization system to calculate the optimal hovering height h* and ejection parameters (v,θ), and generate control command C(t).

[0137] Specifically, deep reinforcement learning algorithms are used to calculate the optimal hovering height in the range of 0.5-2.5m in real time, construct a three-objective optimization framework for seed accuracy, anti-displacement and efficiency, dynamically calculate the optimal seeding parameters, establish the mapping relationship between air pressure and ejection speed and angle, design a state-space feedback controller, and realize real-time closed-loop control of parameters.

[0138] Step S5: Receive control command C(t) through variable stiffness catapult actuator, adjust catapult stiffness k* and seed wrapping parameters, execute seed catapult, and collect actual trajectory data Ta(s) as feedback data input to the adaptive seeding parameter intelligent optimization system for continuous optimization of decision-making strategy.

[0139] Specifically, the spring stiffness is dynamically adjusted by regulating the magnetic field strength, a wind-resistant microstructure array is constructed on the seed surface, an environmentally responsive hydrogel coating layer is provided for the seed, the aerodynamic shape of the seed is optimized, seed ejection is performed, actual trajectory data is collected, the deviation between the theoretical trajectory and the actual trajectory is compared, the correction factor is calculated, the model parameters are updated, and the decision-making strategy is continuously optimized based on reinforcement learning methods.

[0140] Example 11

[0141] The system of this invention was validated in an aerial seeding afforestation experiment on the eastern edge of the Kubuqi Desert in Inner Mongolia. The experimental area covered 500 mu (approximately 33 hectares), with the main vegetation being Artemisia argyi and Salix psammophila, an average wind speed of 8.5 m / s, and the surface mainly consisting of mobile and semi-fixed sand dunes. Two typical afforestation seeds, Pinus sylvestris and Caragana korshinskii, were selected for testing, with 5000 seeds sown in a single aerial seeding attempt.

[0142] The comparison results between the conventional aerial seeding system (system in Comparative Document 1) and the system of the present invention are as follows:

[0143] 1. Seed displacement resistance: Under wind speed of 8.5 m / s, the average displacement distance of seeds in conventional systems within 24 hours after sowing is 3.6 m, and 47% of the seeds are blown away from their original position by more than 5 m; in the system of this invention, the average displacement distance of seeds within 24 hours after sowing is only 1.3 m, and only 16% of the seeds are moved more than 5 m, and the displacement resistance coefficient is increased by 2.8 times.

[0144] 2. Seed accuracy: The average deviation between the seed landing position and the target position in the conventional system is 2.4m, and only 77% of the seeds fall within the predetermined 3m×3m grid; the average deviation of the system of this invention is 0.9m, and 92% of the seeds fall within the predetermined grid, improving accuracy by 15%.

[0145] 3. Seed survival rate: After 30 days of sowing, the average survival rate of the conventional system was 60%, while the average survival rate of the system of this invention reached 85%, an increase of 25%. The survival rate of Pinus sylvestris increased from 55% to 82%, and the survival rate of Caragana korshinskii increased from 65% to 88%.

[0146] 4. Operational efficiency: Conventional systems cannot operate stably when the wind speed is above 6 m / s, with a daily operating area of ​​about 110 acres; the system of this invention can operate stably when the wind speed is below 12 m / s, with a daily operating area of ​​150 acres, an improvement of 36%.

[0147] 5. System Adaptability: Tested in various terrain conditions, including desert edges, hilly areas, and river valleys, the system of this invention demonstrated excellent adaptability. Particularly in areas with complex terrain, the system can automatically adjust aerial seeding parameters, maintaining high seeding accuracy and seed stability.

[0148] After 16 months of monitoring, the vegetation coverage rate in the aerial seeding area achieved by this invention reached 65%, 18 percentage points higher than that of conventional systems. The vegetation height was more uniform, the distribution more rational, and the wind and sand resistance was significantly enhanced. The system operated stably with a low failure rate, and maintenance costs were reduced by 30% compared to conventional systems.

[0149] This invention relates to a microenvironment prediction system and method for aerial seeding based on sand grain movement simulation. By solving the technical challenge of seed displacement under windy and sandy conditions, it provides an efficient and reliable technical means for vegetation restoration and ecological construction in arid and semi-arid regions. The system has been applied in aerial seeding afforestation projects in areas such as the Kubuqi Desert and the Mu Us Sandy Land, achieving significant ecological and economic benefits.

[0150] This invention is not only applicable to desertification control, but can also be extended to various ecological restoration fields such as grassland restoration, mine reclamation, and soil and water conservation. The system's modular design and standardized interfaces make it easy to integrate with existing aerial seeding equipment, lowering the technical barriers and costs for its widespread application.

[0151] Furthermore, the multi-scale micro-environment perception and intelligent prediction technology and the graph neural network-driven rapid flow field simulation method in this invention can also be extended to meteorological monitoring, environmental assessment, disaster early warning and other fields, and have broad application prospects and industrialization potential.

[0152] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A microenvironment prediction system for aerial seeding based on sand grain motion simulation, characterized in that, include: A multi-source heterogeneous microscale sensing network is used to collect wind field data, sand grain movement data, and surface characteristic data, and to generate a standard environmental state matrix. The graph neural network-driven sand-airflow coupling simulation engine is connected to the multi-source heterogeneous microscale sensing network to receive the standard environmental state matrix and generate flow field prediction data through physical information-enhanced graph neural network calculation. A seed-sand multiphysics field coupled behavior prediction model is connected to the sand-airflow coupled simulation engine driven by the graph neural network. It is used to calculate the seed movement trajectory and seed displacement resistance coefficient based on the flow field prediction data, and generate a displacement risk probability distribution map. An adaptive aerial seeding parameter intelligent optimization system is connected to the seed-sand multi-physics field coupled behavior prediction model. It is used to calculate the optimal hovering height and ejection parameters based on the displacement risk probability distribution map, the seed movement trajectory and the seed displacement resistance coefficient, and generate control commands. The variable stiffness ejection actuator is connected to the adaptive seeding parameter intelligent optimization system. It is used to receive the control command, adjust the ejection stiffness and seed wrapping parameters, execute seed ejection, and collect actual trajectory data as feedback data to send to the adaptive seeding parameter intelligent optimization system.

2. The microenvironment prediction system for aerial seeding based on sand grain motion simulation according to claim 1, characterized in that, The multi-source heterogeneous microscale sensing network includes: A miniature PIV particle imaging velocimetry system is used to capture the motion trajectory and velocity field of sand particles at the 10-500μm scale; A miniature six-degree-of-freedom wind field sensor array for measuring the three-dimensional wind field distribution in a 3m³ space; A rapid surface property detector is used to obtain parameters of surface sand layer roughness, density, and water content. The data fusion processing unit is connected to the micro PIV particle image velocimetry system, the micro six-degree-of-freedom wind field sensor array, and the surface characteristic rapid detector. It is used to perform spatiotemporal alignment and fusion of multi-sensor data using the Kalman filter algorithm to generate the standard environmental state matrix.

3. The microenvironment prediction system for aerial seeding based on sand grain motion simulation according to claim 1, characterized in that, The graph neural network-driven sand-airflow coupling simulation engine includes: The Physical Information Augmented Graph Neural Network Module is used to construct a Physical Information Augmented Graph Network structure by using sand grains as nodes and flow fields as edges. A multi-scale hybrid discrete element model is used to combine the discrete element method with computational fluid dynamics to construct a three-phase interactive mathematical model of sand grains-seeds-airflow. The recurrent convolutional GNN accelerator is used to compress traditional CFD computation to within 50ms with the help of pre-trained models; The flow field fast prediction algorithm module is used to predict the movement field of sand particles within a 3m³ space using dynamic adaptive mesh generation technology.

4. The microenvironment prediction system for aerial seeding based on sand grain motion simulation according to claim 1, characterized in that, The seed-sand multiphysics coupling behavior prediction model includes: The multiphysics coupling model module is used to integrate mechanical force fields, fluid force fields and surface energy fields to construct multi-field coupling mathematical models; The seed trajectory prediction algorithm module is used to predict the trajectory of a seed from release to landing based on the Lagrange particle tracking method. The sand grain migration risk assessment engine is used to build a risk assessment model based on Monte Carlo simulation and calculate the displacement probability distribution of seeds under different wind speed conditions after they land. A real-time displacement resistance coefficient calculator is used to construct a seed displacement resistance coefficient prediction model, taking into account seed shape, quality, surface characteristics, and surface parameters.

5. The microenvironment prediction system for aerial seeding based on sand grain motion simulation according to claim 1, characterized in that, The adaptive aerial seeding parameter intelligent optimization system includes: A hovering height dynamic optimizer is used to calculate the optimal hovering height in real time within the range of 0.5–2.5m using deep reinforcement learning algorithms; A multi-objective Bayesian optimization framework is used to construct a three-objective optimization framework for seed accuracy, displacement resistance, and efficiency, and to dynamically calculate the optimal seeding parameters. The air pressure-ejection relationship self-learning module is used to establish the mapping relationship between air pressure and ejection speed and angle; State-space feedback controllers are used to achieve real-time closed-loop control of parameters based on a micro-environment state-space model.

6. The microenvironment prediction system for aerial seeding based on sand grain motion simulation according to claim 1, characterized in that, The variable stiffness ejection mechanism includes: A magnetorheological fluid-based variable stiffness ejection mechanism is used to dynamically adjust the spring stiffness by adjusting the magnetic field strength. A seed surface biomimetic microstructure processing unit is used to construct a wind-resistant microstructure array on the seed surface; An environmentally responsive hydrogel encapsulation device for providing seeds with a composite hydrogel encapsulation layer containing silicates and alginates; The aerodynamic optimization unit is used to optimize the aerodynamic shape of the seed, reduce the wind load ratio, and improve stability.

7. The microenvironment prediction system for aerial seeding based on sand grain motion simulation according to claim 1, characterized in that, The update frequency of the standard environmental state matrix E(t) is no less than 100Hz, the calculation time of the flow field prediction data F(t+Δt) is less than 100ms, and the spatial resolution of the displacement risk probability distribution map P(x,y) is no less than 2cm. 3 .

8. The microenvironment prediction system for aerial seeding based on sand grain motion simulation according to claim 1, characterized in that, The adaptive aerial seeding parameter intelligent optimization system also includes a closed-loop feedback and continuous optimization module, which is used to compare the deviation between the theoretical trajectory and the actual trajectory, calculate the correction factor, update the model parameters, and continuously optimize the decision-making strategy based on reinforcement learning methods.

9. The microenvironment prediction system for aerial seeding based on sand grain motion simulation according to claim 8, characterized in that, The iteration cycle of the closed-loop feedback and continuous optimization module is less than 500ms, and the optimization speed of a single iteration model is no less than 15% / 100 fly-seeds.

10. A method for predicting the microenvironment of aerial seeding based on sand grain motion simulation, employing the aerial seeding microenvironment prediction system based on sand grain motion simulation as described in any one of claims 1-9, characterized in that, Includes the following steps: Wind field data, sand grain movement data, and surface characteristic data are collected through a multi-source heterogeneous microscale sensing network, and a standard environmental state matrix is ​​generated. The standard environmental state matrix is ​​input into a sand-airflow coupling simulation engine driven by a graph neural network, and flow field prediction data is generated by physical information-enhanced graph neural network calculation. Based on the flow field prediction data, the seed movement trajectory and seed displacement resistance coefficient are calculated using the seed-sand multi-physics field coupled behavior prediction model, and a displacement risk probability distribution map is generated. The displacement risk probability distribution map, the seed motion trajectory, and the seed displacement resistance coefficient are input into the adaptive aerial seeding parameter intelligent optimization system to calculate the optimal hovering height and ejection parameters, and generate control commands. The variable stiffness catapult receives the control commands, adjusts the catapult stiffness and seed wrapping parameters, executes seed catapult, and collects actual trajectory data as feedback data to input into the adaptive seeding parameter intelligent optimization system for continuous optimization of decision-making strategies.

Images