Multi-source data driven forest unmanned aerial vehicle arrangement task chain management system and method thereof

Abstract

The invention discloses a multi-source data-driven management system and a multi-source data-driven management method for a task chain of unmanned aerial vehicles in a forest region. The system comprises an unstructured environment probability sensing module which performs semantic completion on under-forest sparse point cloud by using a Transform network and constructs a probability grid; the anisotropic energy consumption analysis module is used for establishing an energy consumption model containing wind field coupling and parasitic power cubic terms; the multi-target task auction distribution module is used for carrying out distributed collaborative distribution based on energy consumption marginal cost by adopting a CBBA algorithm; the global energy efficiency path planning module is used for generating an energy efficiency optimal task chain; the autonomous homing and energy supply module achieves autonomous charging through an inverted-cone-shaped magnetic attraction mechanism. According to the method, the technical problems of lack of forest sensing, rough energy consumption estimation and unreliable field energy supplement are solved, and the sensor layout efficiency and the system robustness are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of forestry automation equipment and environmental monitoring technology, specifically to a multi-source data-driven forest area drone deployment task chain management system and method. Background Technology

[0002] With the rise of the concepts of "smart forestry" and "ecological Internet of Things," high spatiotemporal resolution monitoring of forest ecosystems has become an urgent need for forestry management and scientific research. Monitoring content covers multiple aspects, including forest fire prevention, pest and disease early warning, biodiversity surveys, and carbon sequestration, which typically requires the deployment of a massive number of wireless sensor nodes across large forest areas. Traditional deployment methods mainly rely on manual installation on foot within forest areas, which is not only extremely inefficient and costly, but also poses significant safety risks to personnel in steep terrain or densely vegetated areas.

[0003] In recent years, small multi-rotor unmanned aerial vehicles (UAVs) have been increasingly used in forestry inspections and material delivery due to their advantages such as flexibility, vertical takeoff and landing (VTOL), and low cost. However, when applying UAVs to complex "sensor deployment task chains," existing technologies still face significant bottlenecks in four dimensions: environmental perception, energy management, task scheduling, and autonomous resupply, which limit the realization of their fully autonomous operation capabilities.

[0004] First, there are limitations in environmental perception capabilities. Forest environments are typically high-density, unstructured environments. Traditional LiDAR-based Simultaneous Localization and Mapping (SLAM) algorithms often suffer from severe sparseness or even missing point cloud data in the lower and middle parts of tree trunks and on the ground when flying under forest canopies due to dense foliage (see reference 1). Existing path planning algorithms typically treat these unknown areas simply as "feasible zones" or "unknown zones," lacking the ability to geometrically complete obstacles such as tree trunks and branches. This "perception blind spot" easily leads to collisions between the UAV and undetected small branches or tree trunks on the planned path, or the planning of paths that are actually impassable. Furthermore, existing point cloud completion networks are mostly designed for regular objects (such as tables, chairs, and cars), lacking targeted optimization for trees with specific biological morphological characteristics (such as vertical growth and fractal structures), resulting in significantly reduced completion effectiveness in forestry scenarios.

[0005] Second, the coarseness of energy consumption estimation models. Drones operating in forest areas not only require long-distance flights but also frequent hovering, climbing, and attitude adjustments to deploy sensors. Existing drone task allocation algorithms (such as variations of the Traveling Salesman Problem (TSP) and the Vehicle Routing Problem (VRP)) typically assume, when calculating the cost matrix, that flight energy consumption is proportional to Euclidean distance. This linear assumption is acceptable in ideal, open, windless environments, but its error is significant in the complex micrometeorological environment of forest areas. Studies have shown that wind speed and direction have a decisive impact on the energy consumption of multi-rotor UAVs. Forest wind fields exhibit strong spatiotemporal nonuniformity due to the influence of terrain and vegetation (such as canyon winds and canopy turbulence), and energy consumption when flying against the wind can be several times higher than when flying with the wind. Ignoring the anisotropy of the wind field leads to severely uneven task allocation, often resulting in UAVs running out of power and being forced to land due to headwinds, or wasting transport capacity due to reserving too much safe power.

[0006] Third, the rigidity of task coordination and dynamic allocation. In large-scale sensor deployment tasks, there are often hundreds of task points with different priorities (e.g., fire detection requires priority deployment). Existing centralized allocation algorithms rely on stable air-to-ground communication links; once communication is interrupted (which is common in deep forests), the entire fleet will be paralyzed. Distributed algorithms such as Contract Network Protocol (CNP) enhance robustness, but they often struggle to converge to the global optimum when dealing with multiple coupled constraints (e.g., load limits, power limits, time window limits). While existing consensus packet algorithm (CBBA) performs well in multi-task allocation, its cost function typically does not consider complex nonlinear aerodynamic energy consumption, resulting in allocation results that are not physically optimal.

[0007] Fourth, the challenge of continuous operation and autonomous resupply in the field. Multi-rotor drones typically have a flight time of only 20-30 minutes, making it difficult to cover large areas of forest. Therefore, fully autonomous "hangars" or "charging stations" are essential. However, existing automatic charging technologies are mostly designed for flat urban surfaces (such as contact platforms and wireless charging pads). In forest areas, the ground is often uneven and subject to strong gusts of wind. Traditional planar docking methods are prone to sideslip in strong winds, leading to poor contact and even overturning the drone. There is a lack of a reliable docking mechanism with mechanical self-locking, automatic alignment, and the ability to withstand harsh wind conditions.

[0008] In summary, existing technologies lack a systematic solution that integrates deep environmental semantic perception, refined aerodynamic energy consumption models, distributed collaborative allocation, and highly reliable physical docking. Based on this background, this invention proposes a multi-source data-driven method and system for managing the deployment task chain of unmanned aerial vehicles (UAVs) in forest areas. Summary of the Invention

[0009] This application provides a multi-source data-driven forest area UAV deployment task chain management system and method to solve the technical problems in the prior art, such as lack of perception in unstructured forest environments, large deviation in energy consumption estimation in wind field environments, low efficiency of multi-UAV collaborative deployment, and unreliable field refueling.

[0010] The first aspect of this application provides a drone mission chain management system for deploying sensors in forest areas. The system includes: Unstructured Environment Probabilistic Awareness Module: This module is used to deploy a Transformer-based deep learning network to process sparse point clouds acquired by airborne LiDAR. By introducing relative position encoding and morphological constraints, it performs semantic completion on occluded tree trunks and terrain, generating a voxel map containing the probability distribution of obstacles. Anisotropic Energy Consumption Analysis Module: Based on the Momentum Leaf Element Theory (BEMT) and micrometeorology principles, this module constructs an instantaneous power model for UAVs that includes induced power, surface power, parasitic power, and climb power. In particular, by introducing the cubic term of relative airspeed, it accurately quantifies the nonlinear and anisotropic influence of wind field on energy consumption.

[0011] Multi-objective task bidding and allocation module: Used to run a consensus-based packet algorithm (CBBA) to maximize the sum of task suitability score and network connectivity gain while minimizing marginal energy cost, and to conduct distributed task bidding and allocation among multiple drones.

[0012] Global Energy Efficiency Path Planning Module: Based on the assigned task sequence, under dynamic constraints, mutual exclusion constraints, and dynamic return-to-home energy threshold constraints, it constructs and solves the mixed integer nonlinear programming (MINLP) problem to generate the energy-efficient three-dimensional flight trajectory.

[0013] Autonomous homing and energy replenishment module: Used to utilize energy sensing when a task is completed or when power is critically low. The algorithm plans the return path and controls the drone to land on a magnetic charging base station with an inverted cone-shaped interface, using physical geometry to achieve automatic centering and wind-resistant locking.

[0014] A second aspect of this application provides a multi-source data-driven method for managing the deployment task chain of unmanned aerial vehicles (UAVs) in forest areas. The method includes the following steps: Step S1: Establish a probabilistic raster model of the unstructured forest environment based on semantic completion; collect sparse point cloud data of the forest area using airborne LiDAR and perform voxelization on the sparse point cloud; construct a Transformer encoder network containing relative position encoding and use a self-attention mechanism to calculate the correlation matrix between point cloud features; introduce a morphological consistency constraint loss function to perform semantic completion on the missing tree trunk and ground information, and generate a probabilistic voxel map containing inferred obstacles and passable areas; Step S2: Establish an anisotropic energy consumption model based on aerodynamics and micrometeorology; based on the momentum leaf element theory (BEMT), the instantaneous total power during the UAV flight process is decomposed into induced power for maintaining hover, surface power for overcoming drag, parasitic power for overcoming wind drag, and climb power for overcoming gravity; among them, the parasitic power term is modeled as a cubic function of the modulus of the difference between the UAV ground speed and the local wind speed vector to characterize the anisotropic nonlinear influence of wind direction and intensity on flight energy consumption; Step S3: Establish a multi-target sensor deployment task allocation and dynamic bidding model; define a task revenue function, which includes deployment point suitability score and sensor network algebraic connectivity gain; adopt a consensus-based packet algorithm (CBBA) to maximize the total task revenue minus the marginal energy cost, and conduct distributed task bidding and allocation among multiple UAVs; the marginal energy cost is calculated in real time based on an anisotropic energy consumption model. Step S4: Construct a global energy-efficient optimal deployment path planning problem under multiple constraints; establish a mixed-integer nonlinear programming (MINLP) model with the objective function of maximizing the sum of the total mission benefits and the remaining energy integral of the UAV swarm; the model includes UAV dynamic constraints, mission mutual exclusion constraints, and dynamic return-to-home energy threshold constraints; the dynamic return-to-home energy threshold constraint requires that the remaining energy at any time must be greater than the sum of the energy required for return-to-home against the current wind field and the safety reserve energy. Step S5: Solve the above optimization problem using a hierarchical hybrid strategy and generate control commands; utilize energy sensing. The algorithm searches for the flight corridor with the lowest energy cost in the probabilistic voxel map. The heuristic function includes an energy prediction term based on the wind field. The generated path is smoothed by B-spline to generate the flight trajectory. When the mission ends or low battery is triggered, the drone is controlled to fly towards the inverted cone-shaped charging base station. The passive sliding alignment and magnetic locking are achieved by using the inverted cone interface that meets the mechanical self-locking condition. In step S1, relative position encoding is introduced in the Transformer encoder network. It concerns any two points in a point cloud. Vertical coordinate difference The function is used to enhance the network's ability to perceive the vertical growth topology of trees; the formula for calculating the attention score is:

[0015] in, These are the query matrix, key matrix, and value matrix, respectively. Feature scaling factor; relative position encoding By mapping the vertical distance to attention weight bias through a learnable parameter matrix, voxel features that are adjacent in the vertical direction have higher relevance weights. In step S1, the morphological consistency constraint loss function Defined as the local normal vector of the completion point With gravity vector Sum of squares of dot products:

[0016] Total loss function For chamfer distance loss Loss due to morphological consistency constraints Weighted sum: ,in is the weighting coefficient; this constraint is used to penalize non-vertically growing geometry in the generated point cloud, forcing the trunk point cloud generated by the completion network to conform to the gravity direction characteristics of natural growth, and suppressing the generation of messy noise points; In step S2, the instantaneous total power in the anisotropic energy consumption model Represented as:

[0017] Among them, parasitic power The calculation formula is:

[0018] in, The air drag coefficient of the drone. For air density in forest areas, For the reference area of ​​the drone facing the wind, For the ground velocity vector of the UAV, Local wind speed vector; induced power With drone payload The relationship is:

[0019] in, It is the acceleration due to gravity. The total area swept by all rotors; For blade profile power, For ramp-up power, For the power consumption of airborne electronic equipment; In step S3, the task reward function The calculation formula is:

[0020] in, The suitability score for the placement points is based on topographic slope, soil moisture, and canopy openness. To deploy points Add to existing sensor network diagram Then, the second smallest eigenvalue of the Laplacian matrix of the graph (Fiedler Value) The change in () is used to characterize the degree of enhancement in network communication connectivity; The weighting coefficients are used; the distributed task bidding process employs the CBBA algorithm, and the drone... For the task bid Defined as:

[0021] in, It is a task Insert drone The marginal energy cost added after the current task sequence is obtained by integrating the anisotropic energy consumption model over the inserted path segment; Specifically, in step S4, the dynamic return-to-home battery threshold constraint is as follows:

[0022] in, This assumes the power of the drone returning to the base station along the current shortest path; this power calculation must take into account the real-time wind field vector. If you are currently in a downwind position, you need to fly against the wind to return home. It will increase significantly due to the cubic growth of the parasitic power term, thereby increasing the return-to-home power threshold; In step S5, energy sensing heuristic function of the algorithm Defined from the current node To the target node Minimum estimated energy consumption:

[0023] in, For Euclidean distance, At maximum cruising speed, For height coordinates; The wind field penalty term takes a positive value when the target direction is opposite to the wind direction, and a zero value when they are the same or perpendicular; the heuristic function guides the path search to first extend to the downwind layer or contour line area to avoid the high amount of gravity work generated by the sharp ascent; In step S5, the inverted cone-shaped interface design of the inverted cone-shaped charging base station must meet the following mechanical self-locking and wind resistance conditions: (1) Self-slip condition: To ensure that the UAV relies on the component of gravity in a powerless state after touching the bottom. Overcoming friction Automatically slides into the center electrode, the semi-cone angle of the inverted conical interface coefficient of friction with contact surface Must meet:

[0024] (2) Wind-resistant self-locking condition: To prevent the drone from being blown over by ambient wind during charging, the magnetic attraction mechanism provides magnetic attraction force. Base contact diameter With maximum design wind speed The height of the wind pressure center Torque balance constraints must be met:

[0025] The anisotropic energy consumption model also includes a micro-meteorological wind field reconstruction step: using the difference between the ground speed vector measured by the inertial navigation system (INS) of each UAV in the UAV swarm and the airspeed vector measured by the airspeed tube, the local wind field is inverted in real time. Gaussian process regression (GPR) is used to interpolate sparse sampling points to construct a three-dimensional dynamic wind field map of the forest area, and the wind field map is fed back to the path planning module in step S4 to correct the energy cost weight of each path segment. A management system for performing the method described in any of the preceding items, comprising: The drone swarm and airborne sensing subsystem include multiple rotary-wing drones, each equipped with a lidar, IMU, pitot tube and embedded computing unit, used to collect environmental data and execute flight control commands; The unstructured environment probability perception module is deployed on edge computing nodes or onboard computers to perform step S1, which uses the Transformer network to perform semantic completion and probabilistic raster map construction on the point cloud. The task chain management and energy efficiency optimization module is used to execute steps S2, S3 and S4, maintain the anisotropic energy consumption model, run the distributed auction algorithm to allocate tasks, and solve the global energy efficiency optimal path. The autonomous homing and energy replenishment device includes several inverted cone-shaped magnetic charging base stations distributed in the forest area. The base stations have physical geometric structures that meet the conditions of self-sliding and wind-resistant self-locking, and are used to cooperate with the UAV to perform the autonomous homing and energy replenishment operation in step S5. One or more technical solutions provided in this application have at least the following technical effects or advantages: ●Solvees a fundamental industry problem: the deployment of forestry monitoring equipment relies entirely on manual labor and cannot achieve automated closed-loop systems. This invention integrates four core capabilities: environmental perception, energy efficiency optimization, collaborative scheduling, and autonomous power replenishment. For the first time in the industry, it achieves a fully unmanned closed-loop process from "task assignment" to "sensor deployment." This not only increases deployment efficiency several times compared to traditional manual carrying operations but also extends the operational range to rugged terrain and deep mountain areas inaccessible to humans, significantly reducing the operational risks for forestry workers. It is a revolutionary automated solution for the construction of smart forestry IoT.

[0026] ● This invention addresses the challenges of perception loss and modeling gaps caused by occlusion in high-canopy-density forest environments: By introducing a Transformer semantic completion network that includes vertical relative position encoding, this invention overcomes the limitations of traditional geometric filling algorithms and creatively utilizes the biological prior knowledge of trees' "vertical growth" as a strong constraint. Even with extremely sparse (missing rate >50%) airborne LiDAR point clouds due to canopy occlusion, the system can accurately "fill in" and reconstruct the complete geometric topology of occluded tree trunks. Real-world testing shows that this method improves the completeness of unstructured forest environment modeling from the traditional 60%-70% to over 90%, enabling UAVs to identify and traverse narrow forest gaps, significantly reducing the probability of collisions during autonomous flight, and achieving zero-accident operations in dense forest environments.

[0027] ● This invention solves the problem of severe distortion in energy consumption estimation under complex wind fields caused by traditional linear models, leading to "range anxiety": It abandons the common linear distance energy consumption model (E=k... d) An anisotropic energy consumption model based on the momentum leaf element theory (BEMT) was established. This model accurately reveals and quantifies the nonlinear physical law that the parasitic power of rotary-wing UAVs increases cubically (P∝v³) with relative airspeed, completely eliminating the estimation error caused by "high energy consumption in headwinds". Based on this, the system can calculate the high-precision dynamic return-to-home power threshold in real time, ensuring that sufficient "return tickets" are reserved under any sudden wind conditions, reducing the energy consumption estimation error from 30% in the industry to less than 5%.

[0028] ● This system addresses the problem of ineffective energy dissipation caused by neglecting wind field utilization in traditional path planning: By introducing wind field energy cost maps into global path planning, this system can intelligently identify the flow environment and guide UAVs to actively seek "tailwind corridors" or fly in low-wind-resistance areas sheltered by mountains and canopies, achieving "advantage-seeking and disadvantage-avoiding" energy perception planning. Experiments show that, with the same battery capacity, this wind field-sensitive strategy, which transforms "environmental resistance" into "environmental utilization," can increase the effective operating endurance of UAV swarms by 20%-30% and significantly expand the sensor deployment radius for a single takeoff.

[0029] ● Solves the problems of centralized scheduling failure and multi-machine task conflicts in weak communication environments: Addressing the severe multipath attenuation and frequent communication interruptions of radio signals in forest areas, this invention designs a distributed auction mechanism based on the consensus-based bundle algorithm (CBBA). This mechanism avoids excessive reliance on a central scheduling node, allowing the UAV swarm to achieve consistent task allocation through only local, intermittent communication interactions between neighbors. Even in harsh environments with a communication packet loss rate as high as 25%, it can still guarantee conflict-free convergence of task allocation, effectively preventing resource waste and deadlock caused by multiple UAVs heading to the same target point, greatly improving the system's robustness.

[0030] ● This invention addresses the problems of poor network topology and "data silos" after sensor deployment: It innovatively incorporates network connectivity gain (algebraic connectivity) into the reward function for task allocation. This means that when making decisions, the system not only considers whether the UAV can "reach" the deployment point, but also proactively considers whether the deployment of that node can significantly enhance the overall communication stability of the wireless sensor network (WSN). This unique optimization objective ensures that the monitoring network after deployment not only has wide geographical coverage but also high communication quality, avoiding the deployment of ineffective "dumb nodes."

[0031] ● Solves the problems of low autonomous landing accuracy and unreliable refueling in unstructured terrain: Addressing the challenges of forest environments lacking smooth runways and experiencing gusts of wind, this invention designs an inverted cone-shaped magnetic docking mechanism. This mechanism cleverly utilizes the mechanical balance condition of gravity and friction (tanθ>μ) to achieve passive sliding alignment with immediate correction upon landing deviation. As long as the UAV lands within the cone's area, it will slide 100% into the central contact point. Simultaneously, it utilizes the torque balance condition of strong magnetic attraction at the bottom and wind load (Mmag>Mwind) to achieve wind-resistant locking. This design provides robust physical protection for long-term unattended operations, solving the pain point of unreliable refueling in the field with existing technologies. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0033] Figure 1 The overall architecture and data flow diagram of the forest area UAV deployment task chain management system provided in this application embodiment demonstrate the closed-loop process from multi-source perception, probability modeling, energy consumption analysis to collaborative decision-making and execution.

[0034] Figure 2 This is a schematic diagram of the point cloud semantic completion network structure based on Transformer described in step S1 of this application embodiment, which focuses on the role of relative position encoding and morphological consistency loss function in restoring the vertical structure of the tree trunk.

[0035] Figure 3 The diagram below illustrates the anisotropic power consumption model described in step S2 of this application, showing the aerodynamic characteristic curve of the UAV's flight power changing nonlinearly (cubic order) with relative airspeed (the vector difference between ground speed and wind speed).

[0036] Figure 4 This is a flowchart of the distributed task auction process described in step S3 of this embodiment, illustrating the interactive logic of task bundle construction, marginal cost calculation, and conflict resolution among agents in the CBBA algorithm.

[0037] Figure 5 This is a cross-sectional view of the inverted conical magnetic attraction docking mechanism described in step S5 of this application, and a schematic diagram of the force analysis under a specific transient state. The mechanical equilibrium conditions for achieving gravity-guided sliding and wind-resistant locking are marked in detail.

[0038] Figure 6 The comparison curves of the forest environment perception capability provided in the embodiments of this application with the prior art demonstrate the significant advantages of the method of the present invention in terms of the integrity of environmental modeling compared with traditional LiDAR direct mapping and existing patented technologies (such as CN106200674A) under different leaf area indices (LAI, representing canopy density), especially in the perception stability in high-density forest areas (LAI>4).

[0039] Figure 7 The three-dimensional surface plot comparing the energy efficiency of the anisotropic wind field sensing energy consumption model and the traditional linear model provided in the embodiments of this application demonstrates the improvement in average endurance of the system by optimizing downwind path planning compared to traditional Euclidean distance planning under different wind speeds and wind direction angles.

[0040] Figure 8 The robustness comparison diagram between the distributed task allocation mechanism provided in the embodiments of this application and the prior art demonstrates the ability of the consensus-based allocation algorithm of this invention to maintain the success rate of task allocation compared with traditional centralized scheduling and other similar patents in a weak communication environment where the packet loss rate of communication in forest areas is constantly increasing. Detailed Implementation

[0041] This invention provides a multi-source data-driven method and system for managing the deployment task chain of unmanned aerial vehicles (UAVs) in forest areas. The core of this method lies in deeply integrating prior knowledge of the physical world (tree morphology, aerodynamics, and mechanical principles) into artificial intelligence algorithms (Transformer, CBBA, MINLP) to solve the problem of autonomous operation in complex and unstructured environments.

[0042] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0043] like Figures 1-8 As shown, the system provided by this invention includes five core functional modules, corresponding to steps S1 to S5 in the method flow: 1. Unstructured Environment Probabilistic Awareness Module (corresponding to S1): This module runs on an onboard computing platform (such as NVIDIA Jetson AGX Orin). It receives LiDAR point clouds, performs voxelization, feature encoding, and semantic completion, and finally outputs a probabilistic voxel map. This map not only marks "occupied" and "idle" areas, but also marks "speculated occupied" areas and their confidence levels.

[0044] Anisotropic Energy Consumption Analysis Module (corresponding to S2): This module is the system's "energy auditor." It combines data from the airborne pitot tube, IMU, and weather station to maintain a dynamic energy consumption model in real time. Instead of outputting a single "energy consumption per kilometer," it outputs a value related to the flight vector. scalar field .

[0045] Multi-objective task bidding allocation module (corresponding to S3): This module runs a distributed algorithm. Each drone is an agent, and they exchange task proposals through a local communication network (Ad-hoc). This module is responsible for calculating the compound benefits (suitability + connectivity) and marginal costs of the task, and generating bidding offers.

[0046] Global Energy Efficiency Path Planning Module (corresponding to S4): After determining the task assignment, this module is responsible for generating the specific spatiotemporal trajectory. It solves an optimization problem that includes a time dimension, ensuring that the UAV efficiently completes the task sequence while meeting physical limits and power safety.

[0047] Autonomous homing and energy replenishment module (corresponding to S5): This module includes flight control laws and ground physical facilities. It is responsible for performing high-precision trajectory tracking and executing a special "funnel-shaped" landing procedure when the battery is low, cooperating with the inverted cone-shaped base station to complete the recharging.

[0048] Step S1: Establish a probabilistic raster model of unstructured forest environments based on semantic completion. Traditional SLAM algorithms, when operating in forest canopies, often only capture point clouds containing the treetops and part of the ground, missing the trunks due to canopy shading. This leads the path planner to incorrectly identify the area beneath the canopy as passable. This step uses deep learning to fill in these missing pieces of information.

[0049] Feature Encoding and Attention Mechanism of S11 Sparse Point Cloud. Let the original point cloud be... First, it is voxelized into Input the Transformer encoder.

[0050] The core of Transformer is its self-attention mechanism. This is used to compute voxels. and voxels To assess the correlation between them, we calculate the attention score matrix. :

[0051] in, These are query, key, and value matrices, derived from input features. After linear transformation, we obtain: . It is the square root of the feature dimension, used to normalize the gradient.

[0052] An improvement on S12 Relative Position Encoding (RPE). Ordinary Transformers are insensitive to position or only use absolute position encoding. However, in forestry scenarios, the vertical connectivity of trees is a key feature. That is, if there is a canopy above, roots below, and a trunk in between, there is a high probability. To strengthen this vertical topological relationship, this invention introduces relative position encoding. :

[0053] in, It is a learnable bias matrix that is only related to the relative perpendicular distance between point pairs. Related. When When the distance is small and the horizontal distance is close, It will output larger weights, forcing the network to focus on neighborhood features in the vertical direction, thereby learning to "connect broken trunks".

[0054] S13 Morphological Consistency Constraint Loss. To train the network, we not only need to make the generated point cloud approximate the real point cloud (using Chamfer Distance, CD), but also need to make the generated geometry conform to the biological characteristics of trees (vertical growth).

[0055] Chamfer distance (CD) is defined as:

[0056] Morphological constraints Assuming the tree trunk grows primarily in the opposite direction of gravity. For the completed points We calculate its local normal vector. (Obtained through PCA or neighborhood fitting). If the point belongs to the tree trunk surface, its normal vector should be perpendicular to the growth direction, i.e., perpendicular to the Z-axis. Therefore, we define the penalty term as the sum of the normal vector and the gravity vector. The square of the dot product:

[0057] When the generated point cloud forms a vertical cylinder, the normal vector of the side surface is horizontal and perpendicular to gravity, the dot product is 0, and the loss is minimal; if a messy spherical cloud is generated, the loss will be large.

[0058] Total loss function:

[0059] in It is the balance coefficient, which is usually taken as 0.1-0.5 in experiments.

[0060] Step S2: Establish an anisotropic energy consumption model based on aerodynamics and micrometeorology. Existing algorithms often make simple assumptions This invention, based on rotor aerodynamics, derives anisotropic power equations that incorporate wind field, gravity, and maneuverability characteristics.

[0061] Physical decomposition of S21 instantaneous power. The drone in... Total power at time It consists of four parts:

[0062] (Induced power): Used to generate lift to counteract gravity.

[0063] (Surface power): Overcoming the air viscous drag of the rotor blades themselves.

[0064] (Parasitic power): Overcoming wind resistance generated by the movement of the fuselage in the air.

[0065] (Climbing power): Increases gravitational potential energy.

[0066] (Auxiliary power): Power consumption (constant) of onboard computer and sensors.

[0067] S22 component derivations coupled with wind field. Derivation 1: Induced power. According to the momentum blade element theory (BEMT), during hovering or low-speed flight, the thrust... Induction speed ,in It is air density. It is the sum of the disk areas of all rotor blades. Power Therefore:

[0068] Note: This item displays energy consumption and load capacity. of The power is directly proportional to the power.

[0069] Derivation 2: Parasitic Power (Anisotropic Core). Parasitic power is used to overcome fuselage drag. Drag ,in It is the drag coefficient. It is the windward area. It is relative airspeed.

[0070] Relative airspeed vector ,in It is the speed of the drone relative to the ground. It is the wind speed vector.

[0071] Power of work done against resistance .therefore:

[0072] Key conclusion: Energy consumption is directly proportional to the cube of relative airspeed.

[0073] Headwind conditions: and Reverse direction, relative velocity As the volume increases, energy consumption surges dramatically, reaching cubic levels.

[0074] Tailwind conditions: and When moving in the same direction, the relative speed decreases, resulting in a significant reduction in energy consumption.

[0075] This is the mathematical origin of "anisotropy": at the same ground speed, the energy consumption for flying in different directions is drastically different.

[0076] Derivation 3: Climbing power.

[0077]

[0078] in Vertical velocity. Upgrade. Consumes enormous power; during descent Theoretically, the power is negative (work done by gravity), but in actual motor control, it usually manifests as low power output rather than energy recovery.

[0079] S23 Discretizes the energy consumption cost function. Integrate the above power to the path segment. Up, get Algorithm cost function :

[0080] in , (Only climb costs are calculated).

[0081] Step S3: Establish a multi-target sensor deployment task allocation and dynamic bidding model A market-based distributed approach is used to solve the multi-machine task allocation problem.

[0082] S31 Task Reward Function. For each task to be deployed... Define its benefits :

[0083] Suitability score. Based on the map generated by S1, if... A high score is awarded for a flat ground, an open tree canopy (which facilitates solar charging), and moderate soil moisture.

[0084] Network connectivity gain. Viewing the sensor network as a graph... Calculate the added node Then, the algebraic connectivity of the graph (Fiedler Value) The increase in ). The larger the network size, the stronger its synchronization capabilities and robustness.

[0085] S32 Distributed Bidding (CBBA). Employing the Consensus-Based Bundle Algorithm. Each drone... Maintain a task bundle. and path .

[0086] For a candidate task The drone calculates and inserts it into the current path. Marginal cost after the optimal position:

[0087] Here The function is calculated strictly using the anisotropic model from step S2.

[0088] Bid:

[0089] Logic: If the task Just in the drone On the downwind path, Very small, bid It will be very high, drones This ensures that the task is assigned to the most efficient executor.

[0090] Step S4: Construct the global energy-efficient optimal deployment path planning problem under multiple constraint couplings. After the task assignment is determined, an accurate trajectory needs to be generated. Construct a mixed-integer nonlinear programming (MINLP) model.

[0091] S41 objective function. Maximize the sum of total revenue and remaining energy:

[0092] S42 Key Constraint: Dynamic Return Battery Threshold. The traditional return logic is "remaining battery < 20% return". However, in forest areas, if the return journey is against the wind, 20% battery may not be sufficient. This invention proposes a dynamic threshold:

[0093] in Based on current location and real-time wind field The energy required for return is calculated using the S2 model integral. The system monitors this inequality in real time, and immediately triggers the return abort task once it approaches the critical value.

[0094] Step S5: Solve the above optimization problem using a hierarchical hybrid strategy and generate control commands. The S51 Energy-Aware A algorithm. In grid map search, heuristic functions are used... The term "Euclidean distance" has been changed to "minimum energy estimate".

[0095] This will guide the search tree to avoid towering ridges (which require steep ascents) or areas with strong headwinds, and instead look for "energy troughs" such as valleys or downwind zones.

[0096] S52 autonomous homing: Utilizing the mechanical self-locking principle of an inverted cone-shaped docking system. For outdoor charging, a physical interface of "inverted cone female connector + saucer-shaped female connector" was designed.

[0097] Self-sliding condition: When a drone lands at any point on a cone, it is subject to gravity. and support Function. The component of gravity along the inclined plane. It must be greater than the maximum static friction force .

[0098]

[0099] in It is the half angle of the cone. It is the coefficient of friction of the contact surface. If this condition is met, even if the drone runs out of power, it can automatically slide into the center charging contact under the action of gravity.

[0100] Anti-Wind Locking: The electromagnet engages during charging. Assume the ambient wind speed... The aerodynamic drag torque experienced by the drone is ( (Height of the wind pressure center).

[0101] To prevent it from being blown over, magnetic attraction... The restoring torque generated must be greater than the wind drag torque:

[0102] in This is the contact diameter of the base. This formula guides magnet selection and base size design.

[0103] In summary, this invention achieves efficient, safe, and automated management of UAV sensor deployment tasks in complex forest environments through a full-stack innovation from the perception layer (Transformer completion), the decision layer (CBBA+MINLP), to the physical layer (anisotropic energy consumption and self-locking mechanism).

[0104] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0105] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A multi-source data-driven forest area drone deployment task chain management system, characterized in that, include: The unstructured environment probabilistic perception module is used to perform semantic completion on the collected sparse point cloud using a Transformer network with relative position encoding, and to construct a probabilistic raster map. The anisotropic energy dissipation analysis module is used to establish anisotropic energy dissipation models, modeling flight parasitic power as a cubic function of the vector difference between ground speed and wind speed; The multi-objective task auction allocation module is used to define a reward function that includes deployment suitability and network connectivity gains, and uses the consensus bundle algorithm (CBBA) to allocate tasks with the goal of maximizing net profit. The global energy efficiency path planning module is used to construct a mixed integer nonlinear programming model and solve for the globally energy-efficient optimal path that satisfies the dynamic return threshold by combining dynamic constraints. Autonomous homing and energy replenishment module, used to utilize energy sensing The algorithm generates a trajectory, and the drone is controlled to achieve sliding alignment and locking using an inverted conical magnetic interface.

2. The system as described in claim 1, characterized in that, The unstructured environment probabilistic perception module introduces vertical relative position encoding. The formula for calculating attention score is: ; Construct morphological consistency loss Point cloud completion is achieved by penalizing non-vertical growth structures to constrain point cloud completion.

3. The system as described in claim 1, characterized in that, Parasitic power in anisotropic power consumption analysis module ,in Ground speed and wind speed vectors; instantaneous total power It is decomposed into the sum of induction, profile, climbing and parasitic power.

4. The system as described in claim 1, characterized in that, Task rewards in the multi-objective task bidding and allocation module Bidding Define as profit Subtract the marginal energy cost after inserting the task into the path, which is calculated by integrating the energy consumption model.

5. The system as described in claim 4, characterized in that, List of maintenance tasks and bids for each drone; when receiving a task from a neighbor. bid When this happens, the local machine performs a reset operation, abandoning the task and its subsequent dependent tasks.

6. The system as described in claim 1, characterized in that, The global energy efficiency path planning module sets a dynamic return-to-home power threshold. The integral term represents the energy consumption of returning upwind along the shortest path at the optimal airspeed.

7. The system as described in claim 1, characterized in that, The autonomous homing and energy replenishment module uses energy sensing. Algorithms, Heuristic Functions ,in This is a wind field penalty term, which takes a positive value when the target direction is against the wind.

8. The system as described in claim 1, characterized in that, The inverted cone-shaped interface of the autonomous homing and energy replenishment module satisfies the self-slip condition. ,in It is a semi-cone angle. The coefficient of friction is used to ensure that the drone slides into the center by gravity after touching the bottom.

9. The system as described in claim 8, characterized in that, The interface also meets the wind-resistant self-locking requirements. ,in It is magnetic attraction. For the maximum design wind load, This is the height of the wind pressure center.

10. A method for managing the deployment task chain of unmanned aerial vehicles (UAVs) in forest areas using the system described in any one of claims 1-9, characterized in that, include: S1: Use the unstructured environment probability perception module to perform semantic completion on sparse point clouds and generate a probability raster; S2: Use the anisotropic energy consumption analysis module to calculate the energy consumption cost including the wind field coupling term in real time; S3: Use the multi-objective task bidding allocation module to allocate tasks among multiple machines based on marginal energy cost. S4: Utilize the global energy efficiency path planning module to plan the energy-efficient path while meeting the dynamic return threshold; S5: Utilize the autonomous homing and energy replenishment module to control the UAV to perform tasks and achieve passive glide charging.

Images