A method and system for three-dimensional variable load operation path planning of UAVs for plant protection in mountainous and hilly areas
By employing a path planning method based on multi-stage decoupling and weak coupling of multiple physical quantities, the safety and efficiency issues of UAV operations in mountainous and hilly terrain were resolved. This approach achieved safe and efficient plant protection operations and consistent spraying quality, meeting the requirements for real-time and rapid planning.
Patent Information
- Application Number
- CN202611607386.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-16
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional drone operation path planning methods suffer from problems such as uneven spraying, overlapping or missed spraying, motor overload, and excessive computational load in mountainous and hilly terrain, making it impossible to achieve safe and efficient plant protection operations.
A path planning method with multi-stage decoupling and weak coupling of multiple physical quantities is adopted. Through digital elevation grid extraction, sliding window scanning, multi-dimensional feature clustering, Bayesian optimization and genetic algorithm, a globally optimal three-dimensional variable load operation path is generated. Combined with three-dimensional gravity dynamics and battery safety red line, the UAV can be operated efficiently within the safety boundary.
It improves operational safety and endurance, ensures consistent spraying quality, reduces computational complexity, and meets the needs of real-time and rapid planning.
Smart Images

Figure CN123113945A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of smart agricultural plant protection and low-altitude economy, specifically involving a three-dimensional variable load global scheduling and coverage path planning method and system for plant protection drones based on elevation cyber-physical coupling and multi-stage collaboration, which is suitable for complex terrains such as mountains and hills. Background Technology
[0002] my country has a vast area of mountains and hills, with tea gardens and economic fruit forests mostly distributed in hilly areas with steep slopes. Due to the highly irregular terrain and varying elevations, traditional drone operation path planning methods that rely on manual marking or purely two-dimensional geometric cutting are facing serious practical bottlenecks. First, traditional plot division methods ignore the physical continuity of terrain slope and surface roughness in three-dimensional space. If areas with drastically different terrains are rigidly divided into the same operation plot, the drone will be unable to maintain a constant altitude within the area, leading to uneven atomization of pesticides, overlapping spraying, or serious missed spraying. Second, the liquid pesticide payload carried by agricultural drones usually accounts for more than 40%-50% of the total weight of the drone. During continuous spraying, the payload mass shows a significant non-linear decrease as the spraying is consumed. If the scheduling scheme ignores the dynamic physical coupling between the actual elevation and the variable load, and incorrectly schedules the drone to climb steep slopes at "full load," it will cause the drive motor to experience a sudden overload power spike, which may lead to a mid-air crash due to the momentary undervoltage protection of the battery. Finally, if the scanning angle within a micro-region is directly and forcibly coupled with the timing of cross-operation plot transfers at a macro level, an exponentially increasing nonlinear search space will be formed in three-dimensional space, resulting in excessive computing load on the ground station computer and making it impossible to perform real-time and rapid on-site scheduling. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a three-dimensional variable load path planning method and system for UAVs with phased decoupling and weak coupling of multiple physical quantities. This invention defines multi-dimensional control system state variables and establishes discrete state transition equations for the system. It precisely calculates the multi-dimensional physical constraints between "three-dimensional elevation fluctuations - variable load mass flow - flow sequence," guiding the scheduling system to spontaneously converge to the globally optimal timing sequence: "prioritizing spraying operations in low-altitude or flat areas under heavy load conditions, and then sprinting towards high-altitude areas under light load conditions after significant pesticide consumption and weight reduction." This improves the plant protection area and operational safety per flight while ensuring the actuator's energy saturation safety limit. The specific implementation process of the method is as follows: Step 1, Digital Elevation Grid Extraction and Sliding Window Scanning: Read the GPX / DEM elevation data of the mountainous area to be surveyed and establish... Digital Elevation Matrix Set the sliding window size to (like A full-domain sliding scan sampling is performed on the elevation matrix with a step size of 1. Step 2: Construction of multidimensional terrain physical feature vectors: For each grid window, extract the central row and column spatial coordinates. The average slope is calculated using the first-order central difference operator. ; Calculation of surface roughness using the regional extreme value difference method Construct the original feature vector ; Step 3, Dimensionless transformation and asymmetric weighted matrix calculation: Perform min-max normalization on the eigenvectors to map them to... Space; Introducing a spatial continuity balance weighting factor Calculate the asymmetric weighted eigenvector:
[0004] Step 4, Unsupervised clustering of spatial-physical composite features: Input the global weighted feature matrix into the K-Means algorithm (set the number of cluster centers K), and complete the unsupervised clustering of multidimensional spatial and physical fluctuation features by iteratively minimizing the intra-class variance; Step 5, Convex Hull Mapping and Independent Plot Boundary Cutting: Extract the geometric coordinates of the cluster centers and send them to the Voronoi convex hull generator for spatial geometric mapping. On the DEM base map, lock several independent plant protection operation plots with continuous and seamless boundaries and highly similar internal terrain features.
[0005] Step 6: Generation of multi-angle continuous scanning flight paths: For each independent plant protection operation plot, in continuous angular space... Multiple parallel geodetic sweep lines are generated according to the set angle step size. Step 7: Construction of Variable Load Inertia and Elevation Penalty Evaluation Function: Considering the commutation tail drift, liquid surface sloshing, and climbing work losses under heavy load of the UAV, a local comprehensive energy consumption evaluation function is constructed. :
[0006] In the formula, Number of U-turns This is the gain coefficient. The inertia penalty index for heavy load turning. The initial reagent quality upon entering the work site. This is the sum of the altitude gain during the flight path's ascent. Step 8, Gaussian process surrogate model and UCB adaptive sampling: A Gaussian process (GP) surrogate model is established using a Bayesian optimizer, and an upper confidence bound (UCB) sampling function is used to apply continuous Gaussian perturbation sampling and adaptive iterative optimization to the sweep angle; Step 9, Optimal Angle Locking and 3D Anchor Node Extraction: Converge and output the optimal angle that minimizes the energy consumption assessment value of the work site. The final terrain-following trajectory is generated, and the starting point and entry node of the flight path are precisely located. and the destination leaving node .
[0007] Step 10, Construction of 3D Macroscopic Topology Network: Construct the entry nodes of all plant protection operation plots extracted in Step 2. Leaving the node And the base camp coordinates serve as three-dimensional network nodes for macro-path scheduling; Step 11, Define the state variables of the control system: Define the step size for the UAV to move and schedule across multiple locations. System state vector at time:
[0008] In the formula, For the current network node index, For the remaining drug quality in the fuselage, This represents the current cumulative electrical energy consumption. Step 12, Establishing the Time-Varying Mass Flow State Transition Equations: Constructing the discrete state transition equations of the system:
[0009] If node For spraying segments within the work area, the amount of pesticide deducted will be reduced. ( The route within the plot is long and mimics the terrain. (This refers to the spray flow rate); if it is an empty flight segment for transferring between different work sites, then... ; Step 13, Establishment of a three-dimensional gravity dynamics work and power consumption model: during node transfer and flow time Within this process, the power consumption loss is calculated by integrating the instantaneous power over continuous time. :
[0010] Step 14, One-dimensional chromosome encoding and genetic algorithm topological evolution: Encode the access order of the work plots into an integer chromosome (e.g., Multigenerational genetic evolution is carried out through partial matching crossover (PMX) and swap mutation (Swap) operators; Step 15, Battery Hardware Safety Red Line Saturation Interception and Optimal Sequence Output: Design an adaptive scoring function with boundary saturation interception:
[0011] In the formula, This represents the total power consumption throughout the entire process. This represents the battery's rated safe capacity limit. Exceeding this limit results in immediate rejection, and the final output is the optimal work site access sequence with the lowest total global power consumption and 100% safe return capability.
[0012] The system of the present invention includes: a land parcel division module, configured to: establish a digital elevation model (DEM) matrix of the mountainous terrain to be operated using geographic information data; construct terrain feature vectors by scanning grid by grid through a sliding window; perform maximum and minimum normalization on the feature vectors and introduce spatial and physical asymmetric balance weight factors; apply the K-Means algorithm for unsupervised clustering; and combine Voronoi convex hull mapping to delineate independent plant protection operation parcels that are geographically continuous and have similar internal terrain features. The micro-route optimization module is configured to: generate a continuous terrain-following sweep trajectory in the sweep angle space for each independent work site; construct a local comprehensive energy consumption evaluation function that includes turning inertia penalty and terrain height change penalty; establish a Gaussian process surrogate model using a Bayesian optimizer and perform iterative optimization through the upper confidence bound UCB acquisition function; and output the optimal sweep angle and the entry node In and exit node Out of the route in the three-dimensional geographic space. The global scheduling module is configured to: use the In (entry node), Out (exit node), and base camp coordinates of each work site as topology network nodes; define system state variables including node index, remaining reagent mass, and cumulative power consumption; construct time-varying mass flow state transition equations and a three-dimensional gravity dynamics power consumption model; use the site access order as an integer chromosome and perform genetic evolution using the partial matching crossover PMX operator and the swap mutation operator; and set a rated safe power threshold. The veto interception mechanism outputs the optimal job block access sequence that satisfies the hardware security boundary and has the lowest global total power consumption; The global scheduling module is connected to the UAV flight control system and is used to compile the optimal work site access sequence into a three-dimensional navigation instruction package and send it to the UAV flight control system. The three-dimensional navigation instruction package includes at least absolute latitude and longitude, terrain altitude, flight speed, and the on / off status of spray valves at each node.
[0013] The beneficial technical effects of the present invention are as follows: 1. Significantly improved safety: By establishing three-dimensional gravity dynamics and time-varying mass state equations through steps 12-15, and introducing a battery safety red line to veto interception, the undervoltage crash accident caused by the drone blindly climbing steep slopes under full load and heaviest state is completely eliminated. 2. Intelligent energy saving and spontaneous convergence of timing: After genetic evolution, the system spontaneously converges to the optimal physical timing of "prioritizing operation in low-altitude or flat areas under heavy load and sprinting up steep slopes at high altitudes under light load," thereby increasing the operating area. ; 3. Good spraying quality and particle consistency: The asymmetric clustering in steps 1-5 ensures that the terrain within the work area is highly similar; the Bayesian optimization in steps 6-9 introduces a large inertia reversal penalty of the 1.6th power, which helps to alleviate the problem of pesticide tailing and drift during heavy-load reversal. 4. Fast solution speed: The decoupled control architecture reduces high-dimensional complex optimization to polynomial-level calculation, which can meet the needs of online rapid planning in the field at the second level. Attached Figure Description
[0014] Figure 1 This is a flowchart of the multi-stage collaborative decoupling scheduling control process of the present invention; Figure 2 The images show the results of asymmetric feature weighted clustering and parcel convex hull segmentation in step 1. (a) is three-dimensional, and (b) is two-dimensional. Figure 3 The map shows the sweeping route and three-dimensional entry and exit anchor nodes within the plot based on Bayesian optimization in step 2. (a) is three-dimensional and (b) is two-dimensional. Figure 4 The graphs show the global convergence curves of the genetic algorithm in step 3 under the constraints of dynamic energy consumption and safety red line interception. (a) shows the convergence process of the genetic algorithm, and (b) shows the global scheduling route of the VRP. Detailed Implementation
[0015] The following detailed explanation of the specific steps of this invention, combined with concrete simulation parameters and algorithm calculation processes, is provided. The physical and dynamic calculation parameters of the UAV control system in the embodiment are shown in Table 1.
[0016] Table 1 shows the simulation parameters for the UAV's physical and dynamic aspects. ; In step 1, the ground station reads the map geographic data (GPX or DEM format) of the mountainous area to be operated on, and generates an elevation matrix containing 100×100 sampling grid points. The grid spacing resolution is set to 1.0m. Each grid element... Representing the line, number The three-dimensional elevation (in meters) of the column nodes. The dimensions are set to... An odd-numbered sliding window is used. During scanning, to avoid edge artifacts, padding is applied to expand the original matrix by 2 rows and 2 columns, generating a 104×104 temporary matrix. The sliding window operates with a step size... The system performs a pixel-by-pixel translation scan on the original 100×100 grid to ensure that all 10,000 sampled grid points in the entire area are traversed without omission.
[0017] In step 2, for each sliding window (The center coordinates are located at the ) Okay, number (column), extract and calculate the following four key physical dimensions: Row and column coordinates: Directly extract the index coordinate values of the center grid of the window. , representing a position in a two-dimensional plane; Average slope: Calculated using a first-order central difference operator, the elevation within the window is... Axis (column) and Partial derivatives in the axial (row) direction:
[0018] The geometric slope of the current grid window is calculated as follows:
[0019] Surface roughness (Har): Extracts the highest elevation within a 5×5 window. With the lowest elevation The calculation formula is as follows:
[0020] For each grid point, the above four features are combined to construct the original four-dimensional feature vector:
[0021] For example, based on field data collection, the original feature vector calculated from a very small local grid point is: .
[0022] In step 3, due to spatial coordinates orders of magnitude ( ) is much greater than surface roughness ( Min-Max Normalization must be performed to eliminate the influence of dimensions.
[0023] To enhance the continuity of spatial geography (preventing fragmentation of separated land parcels) while preserving the abrupt changes in topographical physical characteristics, an asymmetric weighting factor is introduced. Weighted eigenvectors The expression is:
[0024] The weighted feature vectors of the 10,000 sampling points across the entire domain are stacked row-wise to generate the global feature matrix. .
[0025] In step 4, the feature matrix is... Input the K-Means clusterer, with the preset number of cluster centers K=3.
[0026] Initialization: Three samples are randomly selected as initial cluster centers using the K-Means algorithm. ; Distance Metrics and Assignment: For Arbitrary Grids Calculate the weighted Euclidean distance between it and each cluster center:
[0027] grid Assign to the nearest cluster ; Centroid Update: Recalculate the mean of all grid feature vectors within each cluster and update the cluster centers. ; Convergence criterion: When the change in centroid between two consecutive iterations... The iteration terminates when the time is reached. After 18 iterations, the algorithm converges and outputs the class label vectors of 10,000 sampled grids. .
[0028] In step 5, the grid cluster data with the same label are extracted, and their spatial coordinates are determined. Mapping back to a two-dimensional geographic plane. Using a Voronoi Diagram & Convex Hull generator, polygon fitting and smoothing of cluster boundaries are performed: the convex hull point set of each cluster is calculated, and stray outliers are excluded (e.g., fragmented micro-operation plots with fewer than 10 points are merged into the nearest adjacent cluster); three seamlessly connected, spatially continuous, and extremely similar independent plant protection operation plots (defined as: operation plot 1 - gentle low-lying plot, operation plot 2 - moderately gentle slope plot, operation plot 3 - steep high-altitude plot) are generated. The final simulation results are as follows: Figure 2 As shown.
[0029] In step 6, in the scanning angular space Within this system, a local coordinate system transformation matrix is established. This is applied to any given sweep angle. :
[0030] Based on the spray width of the agricultural drone =4.0m, in the rotated coordinate system with Make equal-interval parallel straight line cuts at 4.0m intervals, then inversely transform back to the original three-dimensional spatial coordinate system to generate a set of dense parallel geodetic sweep lines. .
[0031] In step 7, for a given route Extract its microscopic geometric features: Total number of turns (Turn_Count): The total number of turning points of the route at the boundary of the work area; Cumulative height of climbing ( ): Extract the elevation of each contour point along the sweep direction, only for the ascending segment ( Accumulate:
[0032] Comprehensive evaluation function combining time-varying load inertia Construction: To prevent frequent operations by drones carrying large quantities of medicine (e.g., 20.0 kg). The strong surface sloshing and motor overload caused by turning necessitate an evaluation equation that includes a 1.6th power inertia penalty term.
[0033] In the formula, The estimated load mass upon entering the work site is temporarily set at 20.0 kg. The 1.6-power exponent ensures that the route with fewer turns under high load conditions has an overwhelming energy consumption advantage.
[0034] In step 8, due to the evaluation function Relying on complex geometric intersection and elevation interpolation, it belongs to a non-explicit black-box function, and Bayesian optimization is used for fast searching: Gaussian Process (GP) Prior Establishment: Constructing a Gaussian process surrogate model, assuming... The kernel function adopts the standard Covariance kernel:
[0035] Initial sampling: in Six angles are randomly and uniformly sampled within the range (e.g.) ), calculate its true Energy consumption value, initialize Gaussian process model; UCB Acquisition Function Adaptive Iteration: Calculating the Upper Confidence Bound (UCB) Acquisition Function:
[0036] In the formula, The mean is predicted by the Gaussian process. This represents the standard deviation (uncertainty) of the forecast. To explore the balance coefficient.
[0037] Optimizer seeks to make The largest next angle After applying a Gaussian perturbation to it, it is then introduced The GP model is calculated and updated. The iteration process is performed only 12 times.
[0038] In step 9, after 12 Bayesian adaptive iterations, the optimal sweep angle that minimizes the overall energy consumption evaluation value is converged and output. (It has been verified that this angle perfectly aligns with the contour lines of the mountain terrain, greatly reducing the work required for climbing and the frequency of turning around).
[0039] Sure Then, the final 3D flight trajectory for the working site 1 is generated. The start and end points of this flight path are extracted and anchored as the 3D node coordinates of the working site: Enter node:
[0040] Leaving the node:
[0041] Similarly, repeat the steps for work plots 2 and 3. The optimal sweep angle is locked and the corresponding 3D entry / exit anchor nodes are extracted: Work site 2: Optimal angle ; Work site 3: Optimal angle Simulation results are as follows Figure 3 As shown In step 10, a macroscopic flow scheduling topology network is constructed, which contains 7 core topology nodes: Node 0: Ground Base Camp Take-off and Landing Point ( ); Node 1: Entry point to work site 1 Node 2: Departure point of work site 1 ; Node 3: Entry point to work site 2 Node 4: Departure point of work site 2 ; Node 5: Entry point to work site 3 Node 6: Departure point of work site 3 .
[0042] In step 11, the system is defined with a discrete scheduling step size. ( Control state vector under: : The index of the network node where the drone is currently located; : The current remaining mass of the agent on the drone body (in kg); From base camp to stride length At that time, the cumulative electrical energy consumed by the system (unit: Wh). Initial state (step size) (The drone is preparing to depart from its base camp.) Set as follows:
[0043] In step 12, the time-varying mass flow state transition equation is established, and the discrete state transition equation of the system is defined as:
[0044] Drug quality deduction The discrete computation logic is as follows:
[0045] In the formula, The physical distance of the spatial trajectory between nodes, with a fixed value per unit spraying rate. .
[0046] In step 13, a three-dimensional gravity dynamics work and power consumption model is established, starting from the nodes. Fly to During the process, the flight time is (speed Because the quality of the pesticide decreases continuously over time during spraying operations within the plot:
[0047] During inter-plot air transport, the quality of the reagent remains constant.
[0048] instantaneous power of the system Disassembly based on basic flight power Power over gravity when climbing hills :
[0049] In the formula, , This is the sine value of the three-dimensional tilt angle. If the drone is flying horizontally or downhill (…), Then the climbing power .
[0050] Integrating the instantaneous power over continuous time (and dividing by 3600 to convert it to watt-hours, Wh) yields the power consumption increment for that state transition. :
[0051] In step 14, the global scheduling optimizer uses a genetic algorithm (GA) to solve for job parcel access timing: Chromosome coding: Integer permutation coding based on work plots is used. Since there are 3 work plots, the chromosome length is 3. For example, chromosome... After the drone takes off from the base camp, it will first execute the operation on operation site 3, then operation site 1, and finally operation site 2. After completing the operation, it will return to the base camp. Crossover operator (PMX): Employs partially-mapped crossover, randomly sets cut points to overlap and exchange gene fragments, and maps and corrects conflicting genes to generate legitimate offspring; Mutation operator (Swap): Employs swap mutation, with mutation probability... Randomly swapping the positions of two genes on a chromosome (e.g.) Mutation ).
[0052] In step 15, chromosomes are evaluated. fitness value When calling steps The state transition equation, from Start by recursively solving the problem segment by segment until you return to the base camp (step size). ), to obtain the total power consumption of the entire process. .
[0053] Set the battery hardware safety power threshold. The fitness calculation logic employs a saturation veto interception mechanism: The final simulation results are as follows Figure 4 As shown.
[0054] In an irrational sequence At that time, the drone, fully loaded with 20.0 kg of pesticide, took off from the base camp and flew directly to the highest and steepest work site, Block 1. Due to... At its maximum value, climbing power achieve Basic power The power consumption reached 470.0W. Recursive calculations show that the power consumption reached 380.0Wh when completing the work on plot 3; when continuing to spray plots 2 and 3 in sequence, the total power consumption quickly exceeded the safety limit. At this point, step 15 triggers a veto mechanism, forcibly setting the fitness of the chromosome to [value missing]. This causes them to be rapidly eliminated in the process of population evolution.
[0055] The system spontaneously converges to the optimal solution when it evolves to the 25th generation. The physical process is as follows: When the drone is fully loaded with 20.0kg, it first operates at the lowest and gentlest work site, Block 2, spraying and consuming 8.0kg of pesticide; then it moves to the medium-slope work site, Block 3, spraying another 7.0kg of pesticide; when it flies to the steepest high-altitude work site, Block 1, only 5.0kg of pesticide remains on the drone (light load sprint). At this point, the power to overcome gravity and climb the slope decreases by 42%, and the total power consumption calculated recursively throughout the process is only 787.77Wh, far below the safety limit of 800.0Wh.
[0056] The algorithm outputs the optimal sequence. The ground station compiles the data into a three-dimensional navigation command package containing absolute latitude and longitude, terrain altitude, flight speed, and the on / off status of spray valves at each node. This package is then transmitted via radio to the UAV's embedded flight control system, enabling highly safe and low-energy autonomous plant protection operations in mountainous areas.
Claims
1. A three-dimensional variable load path planning method for UAVs in mountainous and hilly plant protection, characterized in that, The implementation process of the method includes the following specific steps: Step 1: Utilize geographic information data to establish a digital elevation model (DEM) matrix of the mountainous terrain to be surveyed. ; Set a fixed-size sliding window in Perform a grid-by-grid scan on top; Step 2: For each grid window, extract the center row and column spatial coordinates. Calculate the average slope of the current window. With surface roughness Construct the original four-dimensional terrain feature vector ; Step 3: Perform max-min normalization on the feature vector and introduce a spatial and physical asymmetric balance weighting factor. Calculate the weighted eigenvector ; Step 4: Apply the K-Means algorithm to perform unsupervised clustering on the weighted feature vectors of the entire DEM region, and output a preset number of cluster centers upon convergence. Step 5: Combine the Voronoi convex hull mapping algorithm to map the clustering results back to the geographic space and delineate several independent plant protection operation plots that are geographically continuous and have highly similar internal terrain features. Step 6: For each individual work area, in the sweeping angle space The system generates a continuous, ground-scanning trajectory by setting a step size. Step 7: Construct a local energy consumption evaluation function that comprehensively considers the number of turns, time-varying load inertia drift penalty, and gravity-induced hill-climbing height. ; Step 8: Use a Bayesian optimizer to establish a Gaussian process surrogate model, and perform continuous Gaussian perturbation iterative sampling of the sweep angle based on the upper confidence bound (UCB) acquisition function; Step 9: Output the optimal sweeping angle that minimizes the overall energy consumption assessment value of the work site, and accurately extract the entry node of the route in three-dimensional geographic space. and leaving the node ; Step 10: Enter the nodes of all work plots. and leaving the node And the base camp coordinates serve as the topology network nodes of the macro-scheduling system; Step 11: Define the drone transfer and scheduling step size across multiple locations. System state variables at that time: It includes node index, remaining potion quality, and cumulative power consumption; Step 12: Construct a time-varying mass flow transfer equation for the liquid agent that decreases in real time with the operational flight distance, and calculate the agent consumption transferred between nodes. ; Step 13: Construct a continuous-time integral equation that includes the basic flight power and the power to overcome gravity during hill climb, and calculate the three-dimensional dynamic power loss flowing between nodes. ; Step 14: Encode the integer chromosomes based on the access order of the plots, and carry out multi-generational genetic evolution through partial matching crossover PMX operator and swap mutation Swap operator; Step 15: Set the rated safe power threshold in the fitness function. The saturation veto interception mechanism outputs the optimal job block access sequence that satisfies the hardware security boundary and has the lowest global total power consumption.
2. The UAV three-dimensional variable load path planning method for plant protection in mountainous and hilly areas according to claim 1, characterized in that: In step 3, the formulas for dimensionless and asymmetric weighting are as follows: , ; In the formula, This is the original four-dimensional terrain feature vector; These are the eigenvectors after minimax normalization; and These are the minimum and maximum values of each corresponding feature dimension in the entire sample; A positive term is included to prevent the denominator from being zero; , , and Row coordinates Column coordinates Average slope and surface roughness The normalized value; As the spatial continuity balance weighting factor, take ; The eigenvectors are asymmetrically weighted; superscript This represents the transpose of a vector.
3. The UAV three-dimensional variable load path planning method for plant protection in mountainous and hilly areas according to claim 1, characterized in that: In step 7, the local integrated energy consumption evaluation function The mathematical expression is: ; In the formula, The current candidate scanning angle; For the scanning angle The corresponding local comprehensive energy consumption assessment value; , and These are the gain coefficients for the number of turns, the variable load inertia penalty, and the climb height, respectively. This represents the total number of U-turns and turns along the corresponding route. The remaining amount of pesticide when the drone enters the current work site; index The inertia penalty index for heavy load turning; This represents the cumulative climb altitude for each ascending segment along the stated route.
4. The UAV three-dimensional variable load path planning method for plant protection in mountainous and hilly areas according to claim 1, characterized in that: In steps 11 and 12, the discrete state transition equations of the system are: ; In the formula, For the first The system state vector of the step; For discrete scheduling step size; For the first The network node index corresponding to the step; For the first The remaining amount of medicine on the drone fuselage during the flight; For the node Transfer to node The quality of the medicine consumed during the process; For the first The cumulative power consumption during the step; For the node Transfer to node The power consumption increment generated during the process.
5. The UAV three-dimensional variable load path planning method for plant protection in mountainous and hilly areas according to claim 1, characterized in that: In step 13, the three-dimensional dynamic power consumption loss The calculation equation is as follows: ; In the formula, This represents the three-dimensional dynamic power consumption loss during the inter-node transfer process; For continuous flight time; The total flight time for the current node transfer segment; for Maintain the basic power level for drone flight at all times; for The additional power required to overcome gravity during ascent.
6. The UAV three-dimensional variable load path planning method for plant protection in mountainous and hilly areas according to claim 1, characterized in that: In step 15, the battery hardware safety red line saturation interception mechanism is expressed as follows: ; In the formula, This is used by a genetic algorithm to evaluate the fitness value of the access sequence of job plots; The total power consumption of the drone throughout the entire process of completing all the work sites and returning to the base camp; This represents the discrete scheduling step size at the end of the entire process; This refers to the battery's rated safe charge threshold.
7. A system for implementing the UAV three-dimensional variable load path planning method for plant protection in mountainous and hilly areas as described in any one of claims 1 to 6, characterized in that, include: The plot division module is configured to: establish a digital elevation model (DEM) matrix of the mountainous terrain to be operated using geographic information data; construct terrain feature vectors by scanning grid by grid through a sliding window; perform maximum and minimum normalization on the feature vectors and introduce spatial and physical asymmetric balance weight factors; apply the K-Means algorithm for unsupervised clustering; and combine Voronoi convex hull mapping to delineate independent plant protection operation plots that are geographically continuous and have similar internal terrain features. The micro-route optimization module is configured to: generate a continuous terrain-following sweep trajectory in the sweep angle space for each independent work site; construct a local comprehensive energy consumption evaluation function that includes turning inertia penalty and terrain height change penalty; establish a Gaussian process surrogate model using a Bayesian optimizer and perform iterative optimization through the upper confidence bound UCB acquisition function; and output the optimal sweep angle and the entry node In and exit node Out of the route in the three-dimensional geographic space. The global scheduling module is configured to: use the In (entry node), Out (exit node), and base camp coordinates of each work site as topology network nodes; define system state variables including node index, remaining reagent mass, and cumulative power consumption; construct time-varying mass flow state transition equations and a three-dimensional gravity dynamics power consumption model; use the site access order as an integer chromosome and perform genetic evolution using the partial matching crossover PMX operator and the swap mutation operator; and set a rated safe power threshold. The veto interception mechanism outputs the optimal job block access sequence that satisfies the hardware security boundary and has the lowest global total power consumption; The global scheduling module is connected to the UAV flight control system and is used to compile the optimal work site access sequence into a three-dimensional navigation instruction package and send it to the UAV flight control system. The three-dimensional navigation instruction package includes at least absolute latitude and longitude, terrain altitude, flight speed, and the on / off status of spray valves at each node.
