Agricultural unmanned cluster path planning method based on deep learning

Abstract

The invention discloses an agricultural unmanned cluster path planning method based on deep learning, and relates to the technical field of agricultural mechanization and intellectualization, and the method comprises the following steps: constructing a causal graph neural network comprising environmental parameters, agronomic rule parameters and equipment dynamics parameters; according to the method, dynamic constraint deep fusion of the unmanned aerial vehicle and the unmanned aerial vehicle is realized by constructing a causal association model, and the problem of state space incompatibility of a traditional isomorphic algorithm is solved; by means of a cross-scale causal attention mechanism and a semantic mask technology, field-level data and plant-level agronomic rules are accurately mapped, and a multi-source data semantic fault is broken through; through terrain adaptive modeling and a lightweight model strategy, the complex environment adaptability and the equipment reasoning efficiency are improved, the core problems of low agricultural unmanned cluster cooperation efficiency, poor agricultural compliance, insufficient communication robustness and the like are systematically solved, and an efficient and reliable scheme is provided for global fine operation.

Description

Technical Field

[0001] The present invention relates to the field of agricultural mechanization and intelligent technology, and specifically to a deep learning-based agricultural unmanned cluster path planning method. Background Art

[0002] Agricultural unmanned swarms can achieve full coverage of multiple tasks such as sowing, plant protection, and inspections through the collaboration of drones (capable of three-dimensional spatial operations) and unmanned vehicles (for heavy-duty ground operations). However, existing path planning technology is limited by two core bottlenecks:

[0003] (I) The coupling effect of differences in the dynamic models of heterogeneous devices and communication delays: UAVs are subject to three-dimensional kinematic constraints (Euler angle attitude control, thrust vector dynamics) and airspace regulations, and their state space includes multi-dimensional parameters such as position (x, y, z) and attitude (roll / pitch / yaw angles). Unmanned vehicles are subject to two-dimensional nonholonomic constraints (curvature radius restrictions under the Ackerman steering model) and terrain dynamics constraints (soil friction coefficient, maximum climbing angle), and their state space is centered on position (x, y), heading angle, and drive torque. The difference in their kinematic models invalidates the unified state space assumption of traditional homogeneous multi-robot collaborative algorithms (such as contract network protocols and consensus algorithms).

[0004] In addition, the non-ideal characteristics of communication in agricultural operating environments (such as significant time-varying delay variance and high packet loss rate) and the heterogeneity of equipment communication modules (drone data radio vs. unmanned vehicle 4G / 5G) cause the centralized control architecture to accumulate prediction errors due to state information lag, and the distributed consistency protocol to fail asynchronous adaptation due to dynamic changes in network topology.

[0005] (II) Causal Logic and Cross-Scale Semantic Faults in Multi-Source Heterogeneous Data Fusion: When processing agricultural multi-source data (satellite remote sensing, field sensors, equipment status, agronomic rules), existing technologies only achieve alignment at the spatiotemporal correlation level (such as sensor calibration through TF coordinate conversion in the ROS framework), but fail to establish an agronomic causal reasoning path. Specific flaws are as follows:

[0006] 1. Lack of causal modeling: Data fusion in industrial scenarios relies on correlation analysis such as spatial registration of LiDAR and vision, while agricultural operations require domain-specific causal chains. For example, "soil moisture content <30% + image of curled corn leaves → trigger irrigation path planning." This decision-making requires integration with crop physiological models (such as the Penman-Monteith evapotranspiration model to quantify water demand and the Monsi-Saeki canopy model to analyze the impact of light distribution). Existing technologies lack a causal reasoning mechanism for the "environmental data → crop demand → operational constraints" chain. Consequently, path planning is based solely on statistical correlations (e.g., "historically high slip rates in high-humidity areas") rather than causal logic (e.g., "friction coefficient < 0.3 in clay soil with moisture content > 80%; dynamic constraints prevent high-speed turns").

[0007] 2. Technical gaps in cross-scale semantic unification: There is a three-order-of-magnitude scale difference between cloud-based satellite remote sensing data (field-scale, 10m resolution) and field sensor data (plant-scale, 10cm resolution). Existing methods cannot accurately map "satellite-monitored areas of abnormally high NDVI yield" to "field sub-plot priority areas for operations that incorporate corn root zone avoidance rules." For example, the corn variety "Zhengdan 958" has a large horizontal root radius, requiring unmanned vehicles to meet specific dynamic constraints for their minimum turning radius in this area. Existing technologies lack the semantic connection between these plant-level agronomic rules and field-level data.

[0008] 3. Lack of dynamic causal conditional reasoning capabilities: When meteorological data (wind speed > level 6) and equipment status (drone remaining battery power <20%) are triggered simultaneously, existing technologies rely on preset priority rules (such as "battery power priority") to make decisions, without deeply coupling agronomic rules (such as "operation efficiency decreases when the drone spray width deviation rate is significant under strong winds, and the mission should be interrupted") and equipment dynamic constraints (such as "the drone's remaining battery power only supports limited flight, and the path back to the charging station needs to avoid the headwind area"). This results in collaborative decision-making deviating from agronomic goals and the physical limits of the equipment.

[0009] 3. Systemic defects of existing technologies:

[0010] The core conflict between industrial multi-robot data fusion frameworks (such as early fusion and late fusion) and agricultural scenarios lies in the fact that the former achieves sensor data registration based on spatiotemporal correlation, while the latter requires agronomic knowledge-driven causal relationship modeling and cross-scale semantic unification. Existing research on single agricultural equipment (such as drone visual navigation, which relies on a purely data-driven end-to-end model) or simple parameter adjustments (such as presetting the speed threshold of an unmanned vehicle based on soil moisture) has failed to address the pain point of "rich data but inefficient decision-making." This is due to the low agronomic compliance of path planning, which is manifested in high seedling crushing rates, high rates of pesticide overlap, and high equipment failure rates due to exceeding terrain dynamics constraints.

[0011] IV. Necessity of the invention:

[0012] The efficient collaboration of agricultural unmanned clusters depends on the complete logical chain of "data perception → agronomic intention understanding → constraint generation → path planning", but the existing technology has a fundamental deficiency in the "agronomic intention understanding" link.

[0013] In view of this, a deep learning-based agricultural unmanned cluster path planning method is provided to overcome the above problems. Summary of the Invention

[0014] The purpose of the present invention is to provide an agricultural unmanned cluster path planning method based on deep learning to solve the problems raised in the above background technology.

[0015] To solve the above technical problems, the present invention provides a deep learning-based agricultural unmanned cluster path planning method, comprising the following steps:

[0016] Construct a causal graph neural network that includes environmental parameters, agronomic rule parameters, and equipment dynamics parameters;

[0017] Environmental parameters include soil moisture content, wind speed, and obstacle point cloud density;

[0018] Agronomic rule parameters include seedling area identification and variety-specific constraints;

[0019] Equipment dynamic parameters include the remaining battery power of the UAV, the driving torque of the UAV, and the minimum turning radius;

[0020] Use causal graph neural networks to model the fusion of dynamic constraints of drones and unmanned vehicles, and generate a preliminary draft of the global path including equipment task allocation;

[0021] The global path draft is locally fine-tuned based on real-time sensor data to generate the final path.

[0022] Furthermore, the three-dimensional kinematic constraint model of the UAV is:

[0023]

[0024] The second-order derivative of the drone's position p(x, y, z), i.e., the three-dimensional acceleration vector; R(θ, φ, ψ): The rotation matrix consisting of the roll angle θ, the pitch angle φ, and the yaw angle ψ, used to convert the thrust vector into acceleration in the inertial coordinate system; u: The drone's thrust vector, including vertical thrust and attitude control torque; m: The drone's body mass; g: The acceleration due to gravity; e z : z-axis unit vector of the inertial coordinate system.

[0025] Furthermore, the two-dimensional nonholonomic constraint model of the unmanned vehicle is:

[0026]

[0027] in: The y-direction velocity component of the unmanned vehicle; v: the longitudinal speed of the unmanned vehicle; θ: the heading angle of the unmanned vehicle; Heading angle change rate; δ: front wheel steering angle; L: unmanned vehicle wheelbase.

[0028] Furthermore, the nodes of the causal graph neural network include environmental layer nodes, agronomic layer nodes and equipment layer nodes. The environmental layer nodes are used to characterize environmental parameters, the agronomic layer nodes are used to characterize agronomic rule parameters, and the equipment layer nodes are used to characterize equipment dynamic parameters. The edges between the nodes represent causal relationships, and the edge weights are determined by fitting the field measured data with the Bingham distribution, and an agronomic regularization term is introduced to enforce compliance with the agronomic rule constraints.

[0029] Furthermore, the method for generating the draft of the global path includes:

[0030] Fusing satellite remote sensing data with crop growth model output to generate a heat map of task priorities;

[0031] The improved DeepRL-Net reinforcement learning framework is adopted, and a cross-scale causal attention mechanism is introduced to associate field-level tasks with plant-level constraints. The cross-scale causal attention mechanism is as follows:

[0032]

[0033] Among them, α i,j : the attention weight of field-level node i and plant-level node j; Q(NDVI i ):field-level query vector, generated by linear transformation of satellite NDVI values; K(r root,j ): Plant-level key vector encoding root protection radius r root Agronomic constraints; d k : key vector dimension.

[0034] Furthermore, local fine adjustment includes:

[0035] Based on the real-time image of the UAV to identify the position of the seedlings, a three-dimensional no-fly ellipsoid is generated, and the UAV track height is corrected by the differentiable semantic mask. The three-dimensional no-fly ellipsoid is:

[0036]

[0037] Among them, (x, y, z): spatial coordinates; (x0, y0): seedling center coordinates; r root: Horizontal protection radius of corn root system; z≤2.0m: UAV no-fly altitude threshold; Based on the real-time soil moisture content of the unmanned vehicle, the minimum turning radius is dynamically expanded through cubic spline interpolation.

[0038] Furthermore, the dynamic adjustment formula of the minimum turning radius is:

[0039] R min =R base ·(1+0.3·Ⅱ w＞80％ );

[0040] Among them, R min : Minimum turning radius after dynamic adjustment; R base : Basic minimum turning radius; Ⅱ w>80％ : indicator function.

[0041] Furthermore, a communication delay robustness enhancement mechanism is also included, which includes:

[0042] Build a distributed causal knowledge base containing device status and agronomic rules on the edge, and cache causal reasoning results on the device side;

[0043] The model is dynamically switched according to the communication delay status. The full-precision causal graph neural network model is run during normal communication. The lightweight causal reasoning model is activated when communication is interrupted, and autonomous decision-making is made based on the local knowledge base.

[0044] Furthermore, typical rules for storing distributed causal knowledge bases include:

[0045] Wind speed>6 level ∧E UAV <15% → Execute the charging station priority strategy, and the device caches the causal reasoning results of the last 10 minutes with an update cycle of 5 seconds.

[0046] Furthermore, the obstacle point cloud density in the environmental parameters comes from the voxelization processing of the LiDAR point cloud with a resolution of 20 cm; the seedling area identification in the agronomic rule parameters is based on YOLOv8n recognition with a height of <25 cm.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. Systematic improvement in the collaborative efficiency of heterogeneous equipment: In view of the essential differences between the three-dimensional kinematic constraints of drones (Euler angle attitude control, thrust vector dynamics) and the two-dimensional non-holonomic constraints of unmanned vehicles (Ackerman steering model, terrain dynamic parameters), a causal association model was constructed that includes environmental parameters (soil moisture content, obstacle point cloud density), agronomic rules (seedling area identification, variety-specific root protection radius), and equipment dynamic parameters (drone remaining battery power, unmanned vehicle minimum turning radius). This achieved the first deep integration of cross-modal constraints.

[0049] Significantly reduced constraint conflict rates: By fitting field data with a Bingham distribution to establish causal edge weights and combining agronomic regularization terms to enforce constraint compliance, the cross-device task allocation conflict rate was reduced from 25% to 7% using traditional methods. This avoids typical constraint failure issues, such as unmanned vehicles being assigned to pass through areas with slopes exceeding 20° (verified by V-REP simulation and field measurements).

[0050] Breakthrough in communication delay robustness: The distributed causal knowledge base on the edge and the dynamic model switching strategy on the device (adaptive switching between the full-precision model and the 15MB lightweight model) increase the communication delay tolerance from 100ms to 500ms. The autonomous operation time in the signal blind spot (accounting for 20%) reaches 25 minutes, and the path compliance remains above 93%, which is 4 times higher than the existing technology, solving the problem of state lag error accumulation in the centralized control architecture.

[0051] 2. Multi-source data-driven agronomic compliance innovation: To address the cross-scale semantic gap between satellite remote sensing (field scale 10m) and field sensors (plant scale 10cm), an improved DeepRL-Net framework was used to introduce a cross-scale causal attention mechanism to establish an end-to-end mapping of "satellite NDVI abnormally high-yield areas → plant-level root protection constraints", breaking through the limitation of the industrial multi-robot fusion framework that relies solely on spatiotemporal correlation.

[0052] Centimeter-level operational accuracy is ensured: A three-dimensional no-fly ellipsoid (radius 35cm, height threshold 2.0m) is generated based on the seedling positions identified by YOLOv8n. The drone's track altitude is dynamically corrected to 2.2m±0.1m using a differentiable semantic mask, reducing the seedling crushing rate from 18% to 1.5% (based on 5,000 infrared image detection samples). When the unmanned vehicle's real-time moisture content is >80%, the minimum turning radius is dynamically extended to 4.5m through cubic spline interpolation. The number of slips in clay areas is reduced by 88.9% (based on IMU attitude data statistics), and the pesticide overlap rate is reduced from 17% to 6.8%, fully complying with the NY / T3034-2016 operational specifications.

[0053] Deep coupling of causal reasoning: Crop physiological knowledge such as the Penman-Monteith evapotranspiration model and the Monsi-Saeki canopy model are embedded in the graph neural network to construct causal chains such as "soil moisture content-friction coefficient-turning radius" and "wind speed-spray width deviation rate-operating efficiency". This enables path planning to shift from relying on historical data statistics to a deep coupling of agronomic rules and dynamic constraints. For example, in scenarios with strong winds (> level 6) and low battery (<15%), the equipment can autonomously choose the "avoid headwind areas + prioritize returning to charging stations" strategy to avoid decision deviations from traditional preset priority rules.

[0054] 3. Improved adaptability to complex agricultural environments: Through terrain dynamics modeling and crop variety-specific constraint extension, strong adaptability to unstructured environments is achieved:

[0055] Terrain-sensitive dynamic adjustment: In hilly rice field scenarios (slope > 15°), the unmanned vehicle automatically reduces its speed to 1.0m / s, and the drone's flight altitude dynamically adjusts with the terrain (elevation difference ±5m). Combined with DEM terrain data, "terrace boundary avoidance" constraints are generated, and the compliance of terrace edge operations reaches 95%, an increase of 35 percentage points compared to traditional methods.

[0056] Universal applicability to multiple crop scenarios: Targeting the characteristics of corn varieties such as "Zhengdan 958" (root extension radius of 35 cm) and rice in the tillering period (root extension of 20 cm), the cross-scale semantic extension technology dynamically adapts the plant-level row spacing constraints (deviation <5 cm) and reduces the rice seedling damage rate from 22% to 2.1%, demonstrating the technical solution's ability to accurately adapt to different crop growth cycles.

[0057] 4. A breakthrough in efficient reasoning under limited computing power: Through lightweight model design and localized storage of causal knowledge, the computing power and battery life bottlenecks of unmanned agricultural equipment have been resolved.

[0058] Efficient reasoning at the edge: A lightweight causal inference model (with a parameter size of 15MB and a knowledge distillation compression rate of 80%) is quickly activated when communication is interrupted, completing local path replanning (such as Z-shaped obstacle avoidance for drones and detours for unmanned vehicles with excessive slopes) within 50ms. This reduces path redundancy by 40% compared to the traditional A* algorithm, meeting the stringent requirement of drones with a flight time of ≤30 minutes.

[0059] The risk of equipment failure has been dramatically reduced: the failure rate of equipment due to constraint violations has dropped from 15 times per 100 hectares to 2 times per 100 hectares (according to vehicle sensor logs). The UAV track offset in a force 6 wind environment has been narrowed from 1.2m to 0.4m (RTK-GPS track comparison), significantly improving operational reliability in complex environments.

[0060] 5. Industrial Application Value of the Technical Solution: This invention constructs a complete logical chain of "data perception - causal reasoning - constraint generation - path planning", which for the first time achieves a technological leap from "geometric space planning" to "agronomic intention understanding" for unmanned agricultural swarms.

[0061] Significantly improved operational efficiency: The coordinated work time for sowing, seedling replacement, and plant protection in a 500-hectare corn field was reduced by 25% compared to traditional methods, and the spatial and temporal conflict rate for task allocation was less than 5%, providing an efficient collaborative paradigm for large-scale agricultural production.

[0062] Deep empowerment of domain knowledge: Convert agronomic knowledge such as the root protection specifications in "Corn Cultivation" and the friction coefficient formula in "Soil-Machine System Dynamics" into computable constraints, forming a dual-drive architecture of "agricultural prior knowledge + deep learning", providing a general technical path for the intelligent upgrade of smart agricultural equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a schematic diagram of a deep learning-based agricultural unmanned cluster path planning method of the present invention. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0065] See also Figure 1 , the present invention provides a technical solution:

[0066] See Figure 1 As shown, an embodiment of an agricultural unmanned cluster path planning method based on deep learning:

[0067] 1. Application scenarios and device configuration:

[0068] This example focuses on a 500-hectare summer corn planting area (seedling stage 10-20 days) in the Huanghuaihai Plain. A heterogeneous cluster consisting of P100Pro drones and LF2204 unmanned vehicles was deployed, along with 200 NB-IoT soil moisture nodes (moisture content accuracy ±2%) and 10 six-element weather stations (wind speed accuracy ±0.2 m / s). This system enables coordinated seeding and seedling replacement operations with plant protection. Equipment parameters strictly adhere to agricultural engineering industry standards, such as GB / T20035-2012, "Specifications for Dynamic Modeling of Agricultural Machinery."

[0069] 2. Implementation steps:

[0070] Example 1: Dynamic Constraint Fusion Based on Causal Graph Neural Network:

[0071] Solve the problem: The impact of differences in dynamic models of heterogeneous devices and communication delay coupling:

[0072] 1. Cross-modal dynamics modeling:

[0073] 1.1. 3D Constraints for UAVs: A kinematic model (Formula 1) including Euler angle differential flatness constraints was established, integrating the airspace rules of the Interim Regulations on the Management of Unmanned Aerial Vehicle Flights (e.g., ≥100m from residential areas):

[0074]

[0075] Among them, the rotation matrix R satisfies the differential flatness condition and supports polynomial trajectory optimization;

[0076] The second-order derivative of the drone position p(x,y,z), that is, the three-dimensional acceleration vector (unit:

[0077] m / s 2 );

[0078] R(θ, φ, ψ): The rotation matrix (dimension: 3×3) consisting of the roll angle θ, the pitch angle φ, and the yaw angle ψ, used to convert the thrust vector to acceleration in the inertial coordinate system;

[0079] u: UAV thrust vector (unit: N), including vertical thrust and attitude control torque;

[0080] m: UAV body mass (unit: kg);

[0081] g: acceleration due to gravity (taken as 9.81m / s 2 );

[0082] e z : The z-axis unit vector of the inertial coordinate system (pointing to the center of the earth).

[0083] 1.2. Two-dimensional constraints for unmanned vehicles: The Ackerman steering model and terrain dynamics parameters (Formula 2) are integrated, and the friction coefficient-water content mapping relationship from Soil Mechanics is introduced:

[0084] (Clay area μ=0.25@w=80%):

[0085] in:

[0086] The y-direction velocity component of the unmanned vehicle;

[0087] v: longitudinal speed of the unmanned vehicle;

[0088] θ: heading angle of the unmanned vehicle;

[0089] Heading angle change rate;

[0090] δ: Front wheel steering angle (unit: rad), constrained by the Ackerman steering model (δ≤δ max =30°);

[0091] L: wheelbase of the unmanned vehicle. In this embodiment, LF2204 is 3.2m.

[0092] 2. Causal Graph Neural Network (CGNN) Construction:

[0093] 2.1、Node definition:

[0094] Environmental layer: soil moisture w, wind speed v w , obstacle point cloud density p (from LiDAR point cloud voxel processing, resolution 20cm);

[0095] Agronomic layer: Seedling area mark S seedling (based on YOLOv8n recognition, height <25cm), variety-specific constraints (root protection radius of "Zhengdan 958", derived from "Corn Cultivation Physiology");

[0096] Device layer: UAV remaining power E UAV 、Unmanned vehicle driving torque T UV , minimum turning radius R min (Dynamic update formula: R min =4.2m+0.3m·Ⅱ w>80％ ).

[0097] 2.2 Edge Weight Learning

[0098] By fitting 200 sets of field measured data through Bingham distribution, the “soil moisture content > 85% → friction coefficient μ < 0.25 → R min =4.5m” causal edge, introducing the agronomic regularization term L agri Enforce compliance with the Corn Root Growth Model constraints.

[0099] The cross-device constraint conflict rate was reduced from 25% in traditional methods to 7% (the probability of an unmanned vehicle being assigned to pass through an area with a slope greater than 20° was reduced by 60%); the communication delay tolerance was increased from 100ms to 500ms, and the device state synchronization error was less than 5cm (RTK-GPS measurement).

[0100] Example 2: A hierarchical planning framework for cross-scale causal semantics:

[0101] Solve the problem: Lack of causal logic in multi-source data and cross-scale semantic faults:

[0102] 1. Global rough planning at field scale (resolution 5m):

[0103] 1.1. Input layer: The Sentinel-2 satellite NDVI data (10m resolution) is integrated with the "seedling sparse area" output by the DSSAT model. The task priority heat map is generated through Kriging interpolation, and the "high-yield area operation priority coefficient P = 1.5" is defined;

[0104] 1.2. Model Layer: The improved DeepRL-Net framework introduces a cross-scale causal attention mechanism (Formula 3) to achieve semantic association between the satellite-level "high-yield area" and the plant-level "turning radius ≥ 4.2m":

[0105]

[0106] Among them, the query vector Q encodes the field-level NDVI value, and the key-value pair K / V stores the plant-level root constraints;

[0107] α i,j : the attention weights of field-level node i and plant-level node j;

[0108] Q(NDVI i ): field-level query vector, generated by linear transformation of satellite NDVI values;

[0109] K(r root,j ): Plant-level key vector encoding root protection radius r root etc. agronomic constraints;

[0110] d k : Key vector dimension (set to 128 in this example), used to scale dot products to stabilize training.

[0111] 1.3. Output layer: Generate a draft of the global path including equipment task allocation, ensuring that the spatiotemporal conflict rate between the UAV inspection track and the unmanned vehicle seeding path is less than 5% (based on V-REP simulation verification).

[0112] 2. Local fine adjustment of plant scale (resolution 10cm):

[0113] 2.1. UAV fine-tuning: Based on the seedling positions identified by YOLOv8n, a 3D no-fly ellipsoid with a radius of 35 cm is generated (Formula 4). The track altitude is corrected to 2.2 m ± 0.1 m using a differentiable semantic mask:

[0114]

[0115] in,

[0116] (x, y, z): spatial coordinates;

[0117] (x0, y0): coordinates of the center of the seedling (determined by YOLOv8n detection results);

[0118] r root : Horizontal protection radius of corn root system (35 cm for "Zhengdan 958" in this embodiment);

[0119] z≤2.0m: The drone’s prohibited altitude to avoid crushing seedlings.

[0120] 2.2. Unmanned vehicle fine-tuning: When the real-time moisture content (w>80%) is reached, the minimum turning radius (Formula 5) is dynamically extended through cubic spline interpolation to avoid slipping in clay areas:

[0121] R min =R base ·(1+0.3·Ⅱ w＞80％ ) (Formula 5);

[0122] in,

[0123] R min : Minimum turning radius after dynamic adjustment;

[0124] R base : Basic minimum turning radius (4.2m for LF2204 in this embodiment);

[0125] Ⅱ w＞80％ : indicator function, which takes the value 1 when the soil moisture content \(w>80\%\), otherwise it takes the value 0;

[0126] 0.3: Moisture content sensitivity coefficient, obtained through field test fitting, used to enlarge the turning radius to avoid slipping.

[0127] The seedling crushing rate was reduced from 18% to 1.5% (5,000 samples were detected by drone infrared imaging); the pesticide overlapping spraying rate was reduced from 17% to 6.8%, which is in line with NY / T3034-2016 "Technical Specifications for Agricultural UAV Spraying Operations".

[0128] Example 3: Communication delay robustness enhancement mechanism:

[0129] Solve the problem: Collaboration failure caused by non-ideal communication:

[0130] 1. Construction of distributed causal knowledge base

[0131] 1.1. Edge: NVIDIA Jetson AGX Orin server synchronizes device status and agronomic rules in real time, building a SQLite knowledge base containing more than 5,000 causal rules. Typical rules include:

[0132] Wind speed>6 level ∧E UAV <15% → Execute the "charging station priority" strategy (weight 0.8);

[0133] 1.2. Device side: Cache the causal reasoning results of the last 10 minutes, including "the maximum allowed speed in the clay area is 1.2 m / s" and "the speed limit for unmanned vehicles in the seedling area is 1.5 m / s", with an update cycle of 5 seconds.

[0134] 2. Dynamic model switching strategy:

[0135] 2.1. Normal communication (latency < 150ms): Run the full-precision CGNN model (parameter size 80MB), optimize the global path every 15 seconds, and support real-time coordination of 3D and 2D paths;

[0136] 2.2. Communication interruption (delay ≥ 500ms): Activate the lightweight model (parameter size 15MB, knowledge distillation compression rate 80%) and make autonomous decisions based on the local knowledge base:

[0137] When the drone detects an obstacle, it uses the "root protection distance 35cm" function to generate a "Z"-shaped detour trajectory, replacing the path redundancy of the traditional A* algorithm.

[0138] When the unmanned vehicle detects a slope greater than 18° through the IMU, it automatically selects a safe path with a curvature radius ≥ 5m, combined with the constraint of "maximum climbing angle 20° in clay areas".

[0139] The autonomous operation time in signal blind spots (accounting for 20%) reaches 25 minutes, and the path compliance remains above 93%, which is 4 times higher than the existing technology;

[0140] When communication is interrupted, the equipment failure frequency drops from 15 times per 100 hectares to 2 times per 100 hectares (IMU and LiDAR failure log statistics).

[0141] III. Extended Example: Adaptation of Hilly Terrain to Rice Planting Scenarios

[0142] Application scenario: 200 hectares of hilly rice fields in the Sichuan Basin (tillering stage, row spacing 30 cm):

[0143] Equipment adjustment:

[0144] Drone: DJI Matrice 350RTK (supports terrain following, minimum flight altitude 1.5m);

[0145] Unmanned vehicle: Kubota M704K-C unmanned rice transplanter (minimum turning radius 2.8m, wheelbase 2.5m).

[0146] Targeted improvements:

[0147] 1. Terrain adaptive dynamic modeling:

[0148] Slope-sensitive constraints are introduced: the unmanned vehicle automatically reduces its speed to 1.0 m / s when the slope is greater than 15°, and the drone's flight altitude is dynamically adjusted to the terrain (elevation difference ±5 m);

[0149] A causal chain of "field slope-soil moisture content-friction coefficient" was constructed, and the drivable area was estimated in real time through the XGBoost model.

[0150] 2. Cross-scale semantic extension:

[0151] Satellite data was fused with DEM terrain data (resolution 5m) to generate “terrace boundary avoidance” constraints;

[0152] A new plant-level constraint has been added: "The horizontal extension of the rice root system is 20 cm during the tillering period." The unmanned vehicle path must maintain a deviation of less than 5 cm from the spacing between plant rows.

[0153] Verify data:

[0154] Path planning time in hilly areas is 30% shorter than traditional methods, and compliance for terrace edge operations reaches 95%;

[0155] The rice seedling damage rate dropped from 22% to 2.1%, proving the universality of the technical solution.

[0156] IV. Quantitative verification of technical effects (as shown in the following table):

[0157]

[0158] Summarize:

[0159] This embodiment systematically addresses the issues of heterogeneous device dynamics differences, lack of causal reasoning, and cross-scale semantic discontinuities through a deep learning architecture enhanced by causal knowledge:

[0160] The depth of causal modeling: Existing technologies only process the physical parameters of equipment. This embodiment, for the first time, embeds the root growth model of "Corn Cultivation" and the friction coefficient formula of "Soil-Machine System Dynamics" into the graph neural network to construct a ternary causal relationship of "environment-agronomy-equipment", breaking through the general framework of industrial multi-robot collaboration.

[0161] End-to-end mapping of cross-scale semantics: Different from the grid scaling of existing technologies, this embodiment realizes the semantic transmission from satellite NDVI values to plant-level turning radius through a cross-scale causal attention mechanism. The semantic mask is directly involved in the path generation gradient calculation to ensure centimeter-level agronomic compliance and solve the problem of scale differences of three orders of magnitude.

[0162] Causal reasoning for communication robustness: Compared to the state synchronization mechanism of existing technologies, this embodiment builds a local knowledge base that contains the coupling relationship between agronomic rules and dynamic parameters. This enables the device to make autonomous decisions based on the "soil moisture content-friction coefficient-turning radius" causal chain in an offline state, breaking through the traditional method's reliance on real-time communication.

Claims

1. A deep learning-based agricultural unmanned cluster path planning method, characterized by: The following steps are involved: Construct a causal graph neural network that includes environmental parameters, agronomic rule parameters, and equipment dynamics parameters; Environmental parameters include soil moisture content, wind speed, and obstacle point cloud density; Agronomic rule parameters include seedling area identification and variety-specific constraints; Equipment dynamic parameters include the remaining battery power of the UAV, the driving torque of the UAV, and the minimum turning radius; Use causal graph neural networks to model the fusion of dynamic constraints of drones and unmanned vehicles, and generate a preliminary draft of the global path including equipment task allocation; The global path draft is locally fine-tuned based on real-time sensor data to generate the final path.

2. The agricultural unmanned cluster path planning method based on deep learning according to claim 1, characterized in that: The three-dimensional kinematic constraint model of the UAV is: The second-order derivative of the drone's position p(x, y, z), i.e., the three-dimensional acceleration vector; R(θ, φ, ψ): The rotation matrix consisting of the roll angle θ, the pitch angle φ, and the yaw angle ψ, used to convert the thrust vector to acceleration in the inertial coordinate system; u: The drone's thrust vector, which includes the vertical thrust and attitude control torque; m: UAV body mass; g: acceleration due to gravity; e z : z-axis unit vector of the inertial coordinate system.

3. The agricultural unmanned cluster path planning method based on deep learning according to claim 1, characterized in that: The two-dimensional nonholonomic constraint model of the unmanned vehicle is: in: The y-direction velocity component of the unmanned vehicle; v: the longitudinal speed of the unmanned vehicle; θ: the heading angle of the unmanned vehicle; Heading angle change rate; δ: front wheel steering angle; L: unmanned vehicle wheelbase.

4. The agricultural unmanned cluster path planning method based on deep learning according to claim 1, characterized in that: The nodes of the causal graph neural network include environmental layer nodes, agronomic layer nodes and equipment layer nodes. The environmental layer nodes are used to represent environmental parameters, the agronomic layer nodes are used to represent agronomic rule parameters, and the equipment layer nodes are used to represent equipment dynamic parameters. The edges between nodes represent causal relationships, and the edge weights are determined by fitting the Bingham distribution to the field measured data. An agronomic regularization term is introduced to enforce compliance with the agronomic rule constraints.

5. The agricultural unmanned cluster path planning method based on deep learning according to claim 1, characterized in that: The method for generating the first draft of the global path includes: Fusing satellite remote sensing data with crop growth model output to generate a heat map of task priorities; The improved DeepRL-Net reinforcement learning framework is adopted, and a cross-scale causal attention mechanism is introduced to associate field-level tasks with plant-level constraints. The cross-scale causal attention mechanism is as follows: Among them, α i,j : the attention weight of field-level node i and plant-level node j; Q(NDVI i ):field-level query vector, generated by linear transformation of satellite NDVI values; K(r root,j ): Plant-level key vector encoding root protection radius r root Agronomic constraints; d k : key vector dimension.

6. The agricultural unmanned cluster path planning method based on deep learning according to claim 1, characterized in that: Local fine-tuning includes: Based on the real-time image of the UAV to identify the position of the seedlings, a three-dimensional no-fly ellipsoid is generated, and the UAV track height is corrected by the differentiable semantic mask. The three-dimensional no-fly ellipsoid is: Among them, (x, y, z): spatial coordinates; (x0, y0): seedling center coordinates; r root : Horizontal protection radius of corn root system; z≤2.0m: UAV no-fly altitude threshold; Based on the real-time soil moisture content of the unmanned vehicle, the minimum turning radius is dynamically expanded through cubic spline interpolation.

7. The agricultural unmanned cluster path planning method based on deep learning according to claim 6, characterized in that: The dynamic adjustment formula for the minimum turning radius is: Among them, R min : Minimum turning radius after dynamic adjustment; R base : Basic minimum turning radius; Indicator function.

8. The agricultural unmanned cluster path planning method based on deep learning according to claim 1, characterized in that: It also includes communication delay robustness enhancement mechanisms, including: Build a distributed causal knowledge base containing device status and agronomic rules on the edge, and cache causal reasoning results on the device side; The model is dynamically switched according to the communication delay status. The full-precision causal graph neural network model is run during normal communication. The lightweight causal reasoning model is activated when communication is interrupted, and autonomous decision-making is made based on the local knowledge base.

9. The agricultural unmanned cluster path planning method based on deep learning according to claim 8, characterized in that: Typical rules for storing distributed causal knowledge bases include: Wind speed>6 level ∧E UAV <15% → Execute the charging station priority strategy, and cache the causal reasoning results of the last 10 minutes on the device side, with an update cycle of 5 seconds.

10. The agricultural unmanned cluster path planning method based on deep learning according to claim 1, characterized in that: The obstacle point cloud density in the environmental parameters comes from the voxelization processing of the LiDAR point cloud with a resolution of 20 cm. The seedling area identification in the agronomic rule parameters is based on YOLOv8n recognition and has a height of less than 25 cm.

Images