Method and system for adaptive generation of delivery route for complex terrain rural unmanned aerial vehicle

Abstract

This invention relates to an adaptive method and system for generating delivery routes for unmanned aerial vehicles (UAVs) in rural areas with complex terrain, within the field of route planning. It simultaneously acquires digital elevation model data, 3D wind field model data, and the UAV's dynamic characteristic parameters of the delivery area, extracting multi-physics constraint environment feature maps. Subsequently, the constraint terms are mapped to a power demand space, constructing a control margin potential energy tensor representing the power reserve state. A conditional encoder is used for modal fusion to generate a multi-dimensional guiding condition vector. Based on this, an initial trajectory noise sequence is constructed and input into a conditional diffusion model. Inverse denoising iteration is performed in the latent space, and the deviation gradient of the intermediate trajectory is calculated in real time. Finally, saturation correction logic is invoked to smooth the power change rate and obtain the target route. This invention effectively reduces the interference of complex terrain and meteorological wind fields, improving the dynamic adaptability and safety of route generation.

Description

Technical Field

[0001] This invention relates to the field of flight route planning, specifically to an adaptive generation method and system for unmanned aerial vehicle (UAV) delivery routes in rural areas with complex terrain. Background Technology

[0002] With the widespread application of drone technology in rural logistics and delivery, autonomous flight path planning for drones in complex mountainous environments has become crucial for ensuring delivery efficiency and flight safety. In rural delivery scenarios, due to the combined effects of undulating terrain, vegetation cover, and complex micro-meteorological environments, drones need to possess the ability to respond to terrain obstacles and dynamic environmental disturbances in real time in three-dimensional space to ensure the automation and reliability of delivery tasks.

[0003] However, existing planning schemes primarily focus on finding the shortest geometric path to avoid obstacles in digital space, striving to achieve accessibility at the spatial topology level. This planning logic, based solely on spatial geometry, is prone to mismatches between waypoints and the execution capabilities of the flight platform in real physical flight environments. Especially in areas where abrupt terrain changes cause drastic airflow fluctuations, if the generated path fails to reserve physical space to cope with environmental disturbances, the flight platform is highly susceptible to losing its ability to correct course when maintaining the predetermined trajectory due to hardware output reaching physical limits, leading to control failure and potentially causing collisions. This problem between spatial location planning and platform physical execution limitations severely restricts the airworthiness stability of UAVs in extreme terrain environments. Summary of the Invention

[0004] The purpose of this application is to provide an adaptive method and system for generating delivery routes for UAVs in rural areas with complex terrain. This method has the advantages of reducing the interference of complex mountain terrain undulations and variable weather wind fields on the physical feasibility of route planning, improving the dynamic adaptability of delivery routes and the accuracy of power reserve space measurement, thereby enhancing the safety of UAV delivery missions and the reliability of route generation results in complex environments.

[0005] The objective of this application can be achieved through the following technical solution: Firstly, an adaptive method for generating delivery routes for UAVs in complex terrain rural areas, comprising the following steps: Simultaneously acquire digital elevation model data, three-dimensional wind field model data, and dynamic characteristic parameters of the UAV for the delivery area, including preset power limit values; Spatial correlation analysis is performed on the digital elevation model data and the three-dimensional wind field model data to extract environmental feature maps containing multiple physical constraints; By combining the aforementioned dynamic characteristic parameters, each physical constraint term in the environmental feature map is mapped to a preset power demand space to construct a control margin potential energy tensor characterizing the power reserve state of each spatial coordinate point within the delivery area. By using a preset condition encoder, modal fusion is performed on the environmental feature map and the control margin potential energy tensor to obtain a multidimensional guiding condition vector characterizing the spatial distribution law of the power reserve state. Obtain the preset spatial dimension of the target delivery route to be generated, and construct an initial trajectory noise sequence based on the preset spatial dimension; The initial trajectory noise sequence and the multidimensional guiding condition vector are input into a preset conditional diffusion model. The conditional diffusion model is used to perform multi-step reverse denoising iterations in the latent space defined by the conditional diffusion model. In each denoising iteration, an intermediate trajectory is generated, and the deviation gradient of the intermediate trajectory relative to the multidimensional guiding condition vector is calculated. The preset saturation correction logic is invoked, and the power change rate between adjacent waypoints is smoothed based on the deviation gradient and the power limit value to obtain the target delivery route.

[0006] Secondly, the adaptive route generation system for drone delivery in rural areas with complex terrain includes the following modules: An environmental perception module is used to simultaneously acquire digital elevation model data, three-dimensional wind field model data, and dynamic characteristic parameters of the UAV in the delivery area. The dynamic characteristic parameters include preset power limit values. The feature processing module is used to perform spatial correlation analysis on the digital elevation model data and the three-dimensional wind field model data, and extract environmental feature maps containing multiple physical constraint terms. The potential energy modeling module is used to combine the dynamic characteristic parameters to map each physical constraint in the environmental feature map to a preset power demand space, and construct a control margin potential energy tensor that characterizes the power reserve state of each spatial coordinate point in the delivery area. The condition generation module is used to perform modal fusion of the environmental feature map and the control margin potential energy tensor through a preset condition encoder to obtain a multidimensional guiding condition vector characterizing the distribution law of the power reserve state in space. The data initialization module is used to obtain the preset spatial dimension of the target delivery route to be generated, and to construct an initial trajectory noise sequence based on the preset spatial dimension; The spatial computing module is used to input the initial trajectory noise sequence and the multidimensional guiding condition vector into a preset conditional diffusion model, and use the conditional diffusion model to perform multi-step reverse denoising iterations in the latent space defined by the conditional diffusion model. In each denoising iteration, an intermediate trajectory is generated, and the deviation gradient of the intermediate trajectory relative to the multidimensional guiding condition vector is calculated. The trajectory evolution module is used to call the preset saturation correction logic, and according to the deviation gradient and the power limit value, to smooth the power change rate between adjacent waypoints of the intermediate trajectory to obtain the target delivery route.

[0007] Compared with the prior art, the beneficial effects of this application are: 1. In this invention, by mapping each physical constraint term in the environmental feature map to a preset power demand space, a control margin potential energy tensor representing the power reserve state of each spatial coordinate point in the delivery area is constructed. This helps to perceive the physical execution boundary of the UAV under the constraints of complex terrain and dynamic wind field in the delivery route planning process, alleviates the mismatch between spatial accessibility and hardware power execution limit that may exist in traditional geometric path planning, and improves the physical feasibility of delivery tasks in complex environments.

[0008] 2. In this invention, by using multi-dimensional guiding condition vectors to perform multi-step reverse denoising iterations within the latent space defined by the conditional diffusion model, and combining deviation gradient and power limit value with saturation correction logic to smooth the power change rate, the generated delivery route tends to evolve and converge towards a region with higher power reserves, and the route characteristics tend to meet the linear response range limited by the preset attitude response bandwidth threshold. This reduces the risk of control failure caused by power jumps induced by sudden environmental changes, thereby enhancing the airworthiness stability of the generated route. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the steps of the adaptive generation method for UAV delivery routes in complex terrain rural areas as described in this application; Figure 2 This is a schematic diagram of the modules of the adaptive route generation system for unmanned aerial vehicle (UAV) delivery in complex terrain in this application. Detailed Implementation

[0010] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0011] Example 1, like Figure 1 As shown, the adaptive generation method for UAV delivery routes in rural areas with complex terrain includes the following steps: The system simultaneously acquires digital elevation model data, 3D wind field model data, and dynamic characteristic parameters of the UAV for the delivery area, including preset power limit values. The component used to perform the synchronous acquisition step can be an airborne multi-source sensor array combined with an offline high-precision digital map library. Its function is to construct the underlying physical constraint benchmark of the delivery environment and flight platform, providing high-fidelity data input for subsequent modeling.

[0012] Spatial correlation analysis is performed on the digital elevation model data and the three-dimensional wind field model data to extract an environmental feature map containing multiple physical constraints. The environmental feature map refers to composite constraint mapping data that integrates terrain undulations and airflow distribution. Specifically, it can be achieved by using a spatial convolution algorithm to extract multi-layer features from the elevation gradient and wind vector field. Its function is to transform isolated environmental elements into obstacle avoidance criteria with spatial continuity.

[0013] Combining the aforementioned dynamic characteristic parameters, each physical constraint term in the environmental feature map is mapped to a preset power demand space to construct a control margin potential energy tensor characterizing the power reserve state of each spatial coordinate point within the delivery area. The control margin potential energy tensor refers to the physical field tensor reflecting the UAV's ability to cope with disturbances at a specific spatial point. Specifically, it can be achieved by numerically mapping the aerodynamic model and the propulsion system efficiency model in a three-dimensional grid space. Its function is to quantify the dynamic safety boundary within the entire domain.

[0014] By using a preset conditional encoder, modal fusion is performed on the environmental feature map and the control margin potential energy tensor to obtain a multidimensional guiding condition vector that characterizes the distribution law of the power reserve state in space. The multidimensional guiding condition vector is a semantic feature vector used to guide the generation of the route. Specifically, it can be achieved by weighted fusion of environmental features and potential energy tensor through a cross-modal attention mechanism. Its role is to transform complex physical field constraints into conditional embeddings that can be recognized by the diffusion model.

[0015] Obtain the preset spatial dimension of the target delivery route to be generated, and construct an initial trajectory noise sequence based on the preset spatial dimension. The initial trajectory noise sequence refers to the high-entropy state sequence used as the generation benchmark. Specifically, it can be implemented using a Gaussian noise generator that matches the route time step and the number of spatial coordinates. Its role is to provide an initial evolutionary carrier for the diffusion model.

[0016] The initial trajectory noise sequence and the multidimensional guidance condition vector are input into a preset conditional diffusion model. The conditional diffusion model performs multi-step reverse denoising iterations within the latent space defined by the model. In each denoising iteration, an intermediate trajectory is generated, and the deviation gradient of the intermediate trajectory relative to the multidimensional guidance condition vector is calculated. The deviation gradient refers to the directional deviation vector between the intermediate trajectory and the safety dynamic boundary. Specifically, it can be implemented using a gradient backpropagation algorithm based on a scoring function. Its function is to guide the noise sequence to converge towards a high-power reserve and low-risk region in real time.

[0017] The preset saturation correction logic is invoked, and based on the deviation gradient and the power limit value, the power change rate between adjacent waypoints of the intermediate trajectory is smoothed to obtain the target delivery route. The saturation correction logic refers to a constraint correction procedure to prevent the motor output from exceeding physical limits. Specifically, it can be implemented using a nonlinear penalty function combined with a second-derivative smoothing algorithm. Its function is to ensure that the generated route satisfies spatial reachability while possessing physical-level dynamic execution stability.

[0018] The core innovation of this application lies in: by constructing a control margin potential energy tensor and transforming it into a guiding gradient for a diffusion model, environmental geometric obstacle avoidance during UAV delivery is elevated to an airworthiness generation mechanism based on dynamic power redundancy. This mechanism, through multi-step iteration and saturation correction within the latent space, enables the flight path to spontaneously avoid dangerous areas prone to generator overload, such as high-gradient shear wind fields, fundamentally solving the problem of trajectory loss of control caused by neglecting dynamic constraints in traditional geometric planning methods under complex physical environments.

[0019] This application further proposes that the specific steps for simultaneously acquiring digital elevation model data, three-dimensional wind field model data, and dynamic characteristic parameters of the UAV in the delivery area, wherein the dynamic characteristic parameters include preset power limit values, include: Before the delivery task is initiated, the system synchronously acquires the following three types of data through a unified spatiotemporal reference: The digital elevation model data of the delivery area is based on the WGS-84 geographic coordinate system, recording the absolute altitude h(x,y) corresponding to each surface grid point (x,y) in the area, in meters.

[0020] Three-dimensional wind field model data, recording the wind speed vector at three-dimensional spatial coordinates (x, y, z): ,in, , The horizontal component (unit: m / s) The vertical component (unit: m / s) is derived from the output of a mesoscale meteorological simulation model (e.g., the WRF model) or from spatial interpolation of measured data from regional meteorological stations.

[0021] The dynamic characteristics of the drone, including its rated full-load power. (Unit: W), Total mass of airframe (m) (Unit: kg), Total area of ​​rotor disk (Unit: m) 2 Aerodynamic drag coefficient Maximum flight speed (Unit: m / s) and preset power limit value (Unit: W). Among them... The power limit value refers to the maximum safe power that the drone's power system can continuously output under rated load. Specifically, it can be calculated by multiplying the drone's rated full-load power by a preset safety margin factor (such as 0.85).

[0022] The three types of data are aligned to a uniform 30m×30m grid through spatial coordinate interpolation to ensure that each data point corresponds one-to-one in subsequent spatial correlation analysis and to eliminate coordinate matching errors caused by inconsistent grid resolution.

[0023] The proposed solution ensures strict alignment of three types of heterogeneous data—topography, meteorology, and dynamics—at a uniform raster resolution (e.g., 30m × 30m) through synchronous acquisition and high-precision spatial registration. This reduces the incidence of spatial mismatch issues among multi-source heterogeneous data, lowers coordinate matching errors, and provides a spatiotemporally consistent input foundation for subsequent environmental feature map extraction.

[0024] This application further proposes that the specific process of performing spatial correlation analysis on the digital elevation model data and the three-dimensional wind field model data to extract an environmental feature map containing multiple physical constraint terms includes: First, gradient analysis and sliding window statistics are performed on the digital elevation model data to extract altitude gradient and terrain relief as terrain constraints. Altitude gradient refers to the rate of change of surface elevation with spatial location, which can be calculated using the central difference method to determine the first-order partial derivatives of the elevation field in orthogonal directions. Its function is to quantify the steepness of the surface to identify high-terrain risk areas. Terrain relief refers to the elevation range within a local area, which can be calculated using a 5×5 grid sliding window to determine the difference between the maximum and minimum elevation values ​​within the window. Its function is to characterize the ruggedness of the terrain to cover the response and prediction range of the flight control system.

[0025] Secondly, flow field analysis is performed on the 3D wind field model data. Local wind speed vectors and canyon acceleration factors are extracted as meteorological constraints by comparing trilinear interpolation with the channel cross-sectional area. The local wind speed vector refers to the velocity and direction of airflow at a specific spatial point, which can be achieved by trilinear interpolation of the original wind field model grid points. Its function is to provide high-precision airflow dynamic input. The canyon acceleration factor refers to the airflow acceleration ratio caused by the terrain contraction effect. It can be estimated by the ratio of the reference channel cross-sectional area to the local channel cross-sectional area calculated based on terrain constraints. Its function is to quantify the amplification effect of complex terrain on the impact kinetic energy of airflow.

[0026] Finally, the terrain constraints and meteorological constraints are overlaid and mapped using spatial coordinate indexing to generate a multi-channel environmental feature map. The environmental feature map refers to a structured data tensor that integrates multiple source physical constraints. Specifically, it can be implemented by combining multiple constraint channels into a multi-dimensional NumPy array based on spatial coordinates. Its function is to transform fragmented environmental data into a unified physical quantization framework.

[0027] The proposed solution extracts a structured environmental feature map by deeply correlating topography and meteorological elements in multiple dimensions. This process enables a refined characterization of the physical constraints of the delivery area, reducing the incidence of problems arising from the lack of a unified quantitative framework between topography and meteorological constraints in traditional methods. This allows for a more accurate description of the nonlinear interference of complex rural environments on UAV flight.

[0028] This application further proposes that, in conjunction with the aforementioned dynamic characteristic parameters, each physical constraint term in the environmental feature map is mapped to a preset power demand space. The power demand space refers to the energy measurement domain defined by the physical output limits of the UAV's power actuators. Specifically, this can be achieved by converting various environmental physical constraints into a set of values ​​in watts (W), which serves to establish a unified quantitative comparison benchmark between different modal constraints. The specific steps for constructing the control margin potential energy tensor characterizing the power reserve state of each spatial coordinate point within the delivery area include: First, a power consumption model for the UAV is constructed based on its dynamic characteristic parameters. This model unifies the discrete physical terms such as altitude, wind speed vector, and canyon acceleration factor in the environmental feature map into an energy demand space measured in power (watts). The specific mapping path is as follows: The model consists of hovering and lift power. Aerodynamic drag power and disturbance compensation power It consists of three parts.

[0029] (1) Mapping of altitude to lift power space: Using rotor momentum theory, the absolute altitude z in the environmental feature map is mapped to the hovering power required to overcome gravity: ; in, The total mass of the drone is expressed in kilograms (kg). Let be the acceleration due to gravity, and take . ; The rotor's induced speed is given by the equation Sure, The total area of ​​the rotor disk is expressed in square meters. ); Current flight altitude Atmospheric density at that location, in units of Its calculations involve: (Standard atmospheric density at sea level, 1.225) (Vertical temperature lapse rate, 0.0065) (Sea level standard temperature, 288.15°C) (Molar mass of dry air, 0.0289644) (Universal gas constant, 8.314) and (Absolute altitude).

[0030] (2) Mapping of meteorological constraints to drag power space: aerodynamic drag power Considering the relative velocity of the drone with respect to the local airflow The calculation formula is: ; in, The aerodynamic drag coefficient is a dimensionless constant given by the factory parameters. The frontal cross-sectional area of ​​the fuselage is expressed in square meters. ); The velocity vector of the drone relative to the airflow is equal to the ground velocity vector. With wind speed vector difference; This is the flight velocity vector of the UAV relative to the ground; This represents the magnitude of the flight velocity vector.

[0031] (3) Mapping of topographic contraction effect to disturbance compensation power space: disturbance compensation power Based on the canyon acceleration factor in the environmental feature map Estimating the wind speed using the local wind speed modulus, i.e.: ; used to quantify the additional power requirements of the flight control system to resist the kinetic energy of airflow impact.

[0032] in, This is the disturbance compensation coefficient, in units of This reflects the system's compensation requirement for unit airflow disturbance energy; The canyon acceleration factor is used to describe the degree of airflow acceleration caused by terrain contraction (dimensionless). The magnitude of the local wind speed vector reflects the overall magnitude of the airflow. Secondly, using the aforementioned UAV power consumption model, the spatial coordinates of each point in the environmental feature map are calculated. Total estimated power consumption required to maintain stable flight attitude Subsequently, combined with the preset power limit value... Calculate the percentage of remaining reserve power at each coordinate point. The formula for its calculation is: ; in The calculation truncates the coordinates of the points where the power demand exceeds the limit. In a physical sense, this is marked as a power-absolute no-fly zone. Finally, the coordinates of each point in the entire space are... Values ​​indexed by three-dimensional space Arrange and construct shapes as Control margin potential energy tensor .

[0033] This application's solution unifies and quantifies terrain constraints (such as altitude influence) and meteorological constraints (such as wind speed and canyon effect influence) into the power reserve state of each spatial coordinate point through a power consumption model. This enables the perception of the physical execution boundary of the flight platform, mitigating the technical deficiency of traditional geometric path planning methods in predicting power saturation risks. Parameters related to environmental feature maps, such as terrain undulation, local wind speed, and canyon acceleration factor, are crucial for constructing the power consumption model. and The core principle of this project is to transform the complex external environmental field into a potential energy distribution that is highly matched with the UAV's own dynamic response. This helps to enhance the control gain of the UAV in dealing with sudden disturbances in complex rural airflow environments at the physical level, in addition to the geometric obstacle avoidance characteristics of the generated flight path, thus ensuring the safety of the flight process.

[0034] This application further proposes that the specific steps for obtaining a multidimensional guiding condition vector characterizing the spatial distribution law of the power reserve state by performing modal fusion on the environmental feature map and the control margin potential energy tensor through a preset condition encoder include: First, the condition generation module generates an environmental feature map. The spatial feature encoding subnetwork of the input conditional encoder is a neural component that extracts high-dimensional features of the terrain and meteorological patterns of the delivery area. Its role is to capture complex environmental spatial constraints. A three-dimensional convolutional neural network (3D-CNN) is used to extract spatial feature representations. This subnetwork consists of four 3D convolutional blocks, which extract spatial pattern information of terrain undulation and wind field distribution through layer-by-layer convolution. The extraction process follows the formula below: ; in, The input shape is Environmental feature map, , , These represent the number of grid cells in the three spatial dimensions; to These represent the operations of four 3D convolutional blocks, each containing a 3D convolutional layer, a batch normalization (3D-BN) layer, and a ReLU activation function; This represents a global average pooling layer, used to compress the feature map into a 256-dimensional vector.

[0035] Secondly, the control residual potential energy tensor The input physical attribute encoding subnetwork is a neural component that performs semantic extraction of the global distribution pattern of power reserve, and its role is to characterize the spatial gradient features of power margin; the physical attribute representation is extracted using a fully connected (MLP) structure. The extraction process follows the formula below: ; in, For shape The control margin potential energy tensor; This indicates that the tensor is flattened and transformed into a one-dimensional vector. This represents a multilayer perceptron consisting of three fully connected layers. The first two layers contain 512 and 256 neurons respectively and are configured with a Dropout layer with a ratio of 0.1. The third layer outputs a 256-dimensional vector representing physical properties.

[0036] Subsequently, the cross-attention fusion module uses the cross-attention mechanism to calculate and Correlation weight matrix between The cross-attention fusion module refers to a computational architecture that aligns heterogeneous features. Its function is to achieve deep coupling between spatial patterns and physical attributes by calculating association weights. The calculation process is as follows: ; ; in, , , These represent the linear projection matrices of the eigenvectors. , , The resulting sequence of queries, keys, and values ​​obtained through mapping; , , The projection matrix is ​​learnable; The number of attention heads (with a value of 8). The dimension of each attention head (with a value of 32); This is a scaling factor used to prevent the result of the dot product from entering the scaling factor. The saturation region of a function; This is a normalized exponential function used to generate association weights; Finally, output attention (Based on association weight) value sequence Weighted summation and linear projection (obtained) and Perform modal fusion and output a multidimensional guiding condition vector. The multidimensional guiding condition vector refers to the feature embedding that characterizes the distribution law of power reserve across the entire domain. Its role is to provide generation guidance with physical constraints for the diffusion model.

[0037] The calculation process follows the formula below: ; in, This is the output vector of the cross-attention function; It is a position-wise feedforward network containing two fully connected layers, used to perform nonlinear feature transformation; The representation layer is normalized to stabilize the feature distribution; This is the output dimension projection matrix, used to expand the features to 512 dimensions to fit the subsequent diffusion model.

[0038] The proposed solution achieves deep fusion of environmental spatial modes and physical dynamic modes through a conditional encoder. Compared to simple feature overlay, the cross-attention mechanism establishes semantic relationships between terrain undulations, wind field intensity, and power reserves, ensuring that the guidance signal accurately reflects the nonlinear constraints of the delivery environment on the UAV's execution capabilities. It is logically linked to feature extraction and potential tensor construction; the environmental feature map provides the spatial semantic foundation for this step, while the potential tensor provides explicit physical safety input, transforming complex external physical field information into a guidance vector understandable by the conditional diffusion model. This enables the subsequently generated flight path to not only avoid terrain obstacles but also spontaneously converge towards high power reserve areas, thereby improving the UAV's physical airworthiness in complex rural environments while ensuring geometric accessibility.

[0039] This application further proposes that the specific steps for obtaining the preset spatial dimension of the target delivery route to be generated and constructing an initial trajectory noise sequence based on the preset spatial dimension include: First, the data initialization module obtains the starting and ending coordinates of the delivery task and calculates the straight-line distance between them. Subsequently, combined with the preset waypoint spacing... Through formula Determine the total number of waypoints on the target delivery route. Next, based on the total number of waypoints... Construct a preset spatial dimension in the original coordinate space, wherein the preset spatial dimension is: ,correspond 3D coordinates of each waypoint Finally, based on the preset spatial dimension, a standard Gaussian distribution is applied. Perform independent and identically distributed sampling to generate an initial track noise sequence. : ; in, for A unit covariance matrix, representing dimensions that are independent of each other and have uniform variance. The generated As the starting state of the reverse denoising process of the conditional diffusion model, it is passed to the spatial computing module.

[0040] The target delivery route refers to a three-dimensional coordinate sequence arranged in flight order from the delivery origin to the destination. This can be achieved through a discretized mapping of straight-line distance and a preset step size, providing clear spatial trajectory constraints for drone delivery. The initial trajectory noise sequence refers to high-entropy state data that serves as the baseline for the evolution of the generation model. This can be generated by randomly sampling from a standard normal distribution. A random matrix is ​​used to implement this, which serves to provide the random starting point with the greatest uncertainty in the diffusion process. The preset spatial dimension refers to the shape of the tensor representing the geometric scale of the route, which is specifically determined by the number of nodes and coordinate components after the route is discretized. Its function is to define the mathematical boundary for matrix operations performed by the model in the latent space.

[0041] This application's solution constructs an initial high-entropy sequence with standard statistical properties through discretization analysis based on the distance between the delivery origin and destination. This ensures sufficient search space and evolutionary degrees of freedom in the route generation process. Furthermore, a standardized sampling mechanism reduces the dependence of the generated results on specific initial states, contributing to improved global search capabilities for route generation. By transforming the defined geographical span into standardized random noise, the diffusion model can start from an unbiased random state and, guided by multi-dimensional guiding condition vectors, gradually eliminate uncertainties in the latent space, thus giving rise to adaptive routes that avoid complex terrain obstacles while meeting dynamic reserve requirements. This process balances the stochastic exploration and physical determinism of route generation, improving the system's path discovery efficiency in extreme mountainous environments.

[0042] This application further proposes that the initial trajectory noise sequence and the multidimensional guiding condition vector are jointly input into a preset conditional diffusion model, and that multi-step reverse denoising iterations are performed within the latent space defined by the conditional diffusion model. In each denoising iteration, an intermediate trajectory is generated, and the deviation gradient of the intermediate trajectory relative to the multidimensional guiding condition vector is calculated. The specific steps include: First, the spatial computing module uses a preset encoding mapping algorithm. The lower dimension of the original coordinate space is Initial track noise sequence Mapping to the latent space, which refers to the continuous feature domain after encoding and mapping the high-dimensional track coordinate sequence, can be achieved by highly compressing the spatial correlation features through a 4-layer nonlinear fully connected network. Its role is to improve the computational efficiency of denoising iteration and facilitate physical constraint guidance through gradient descent.

[0043] Encoding Mapping Algorithm A 4-layer fully connected coding network structure is adopted, with the following specific parameter configuration: the input dimension of coding layer 1 is... (Number of waypoints multiplied by number of coordinate axes), output Dimension; Coding layer 2 maintenance Dimensionality; Coding layer 3 reduces dimensionality to Dimension; Coding layer 4 output Initial noise representation of the latent space of dimension Each layer of encoding is followed immediately by layer normalization and the GELU activation function to ensure the numerical stability of the high-dimensional mapping. ; in, The initial track noise sequence has a dimension of . The initial high-entropy state of the route is generated by sampling from a standard Gaussian distribution; N is the total number of waypoints on the target delivery route, which is determined by the straight-line distance D between the origin and destination and the preset waypoint spacing. Sure; To flatten the operation, The matrix is ​​transformed into a matrix of length . A one-dimensional vector; The preset encoding mapping algorithm (4-layer fully connected encoding network) is used to map the original coordinate space to the latent space; The initial noise representation of the latent space has a dimension of 512 and serves as the starting state for the reverse denoising iteration.

[0044] Within the hidden space, the system proceeds step by step over time. , , , (This embodiment takes) Perform controlled denoising iterations. Each iteration includes three sub-steps: noise prediction, gradient deviation correction, and state sampling. Noise Prediction: Denoising Backbone Network Employing a Transformer-based temporal network, with the current latent space state... Time step position encoding and multidimensional guiding condition vector Using the input, predict the denoised residual for the current step. , ;in, This represents the time step of the current iteration, and its value range is... , , , (This embodiment takes) ); For a moment The latent space state vector, with dimension . ; For time step encoding, the time generated by sinusoidal position encoding. The high-dimensional embedding vector, with dimension . ; The multidimensional guiding condition vector is a guiding signal generated by the condition encoder, carrying environmental constraints and power reserve laws, with a dimension of [missing information]. ; This is the denoising backbone network used to predict the denoising residuals of the current step; The original noise predicted by the network (denoising direction); The network consists of a 6-layer Transformer encoder, the core of which calculates track features and vectors through a conditional cross-attention module. The correlation ensures that the generation process is guided by the constraints of the power reserve distribution.

[0045] Deviation Gradient-Guided Correction: In each iteration, the system calculates the deviation gradient of the intermediate track relative to the physical safety boundary in real time. The deviation gradient refers to the derivative vector that quantifies the deviation of the intermediate track from the power safety margin boundary. Specifically, it is achieved by automatically differentiating and calculating the gradient of the loss function relative to the latent variables. Its role is to drive the generation process to spontaneously evolve towards a spatial region with low load and high power reserve. The calculation process follows the following logic: Decoding intermediate states: using decoding and reconstruction algorithms hidden space state intermediate track restored to coordinate space : ; Where N is the total number of waypoints on the target delivery route; Reshape refers to the operation of changing the dimensional structure (i.e., shape) of data without altering its content and value. Decoding and reconstruction algorithms refer to the inverse mapping process that regresses abstract features to physical coordinates. Specifically, they are implemented by using a decoding network that is symmetric to the encoder parameters to restore latent variables into a discrete waypoint sequence. Their role is to generate the final path points that can be executed by the UAV.

[0046] Physical Deviation Quantification: Extraction Instantaneous power requirements at each track point and from vector The power reserve distribution characteristic value of the corresponding point is extracted by trilinear interpolation. Subsequently, the physical distance deviation value was quantified. : ; Gradient transformation: Constructing a guiding loss function Calculate its relative to the automatic differential framework The partial derivatives are obtained. dimensional deviation gradient : ; Prediction correction: Apply the deviation gradient to the prediction noise according to Equation 20 to obtain the guided prediction noise. , ; in, The intensity coefficient is used to adjust the influence weight of the physical deviation correction signal. In this embodiment, it is taken as... ; To deviate from the gradient, quantize the derivative vector in latent space of the deviation between the intermediate trajectory and the power safety boundary; To guide the corrected prediction noise; State sampling update: Update the latent space state according to the corrected noise using the DDPM backsampling formula. , ; The noise retention coefficient for the current step. ; For noise dispatch coefficients (from) arrive Linear interpolation); The cumulative noise retention factor. ; The current sampling standard deviation, ; To obtain from the standard Gaussian distribution Random noise sampled in the middle (only in (For use at times) This is the updated latent space state vector for the next step; Finally, the final state of the latent space after iteration. Through a preset decoding and reconstruction algorithm Restore to the original coordinate space. Decoding and reconstruction algorithm. It adopts a 4-layer fully connected architecture symmetrical to the encoding network (output dimensions are as follows). Each layer, combined with layer normalization and the GELU activation function, ultimately outputs... dimensional route candidate coordinate sequence : ; in, This is the final state vector in the latent space after the reverse denoising iteration is completed; The preset decoding and reconstruction algorithm (a 4-layer fully connected network symmetrical to the encoder) is used to restore the latent features to the original coordinate space; For the reshaping procedure, the flattened dimensional vector restored to The trajectory coordinate matrix; The final set of waypoint coordinates is generated from the candidate route sequence through multi-step denoising.

[0047] This application's solution transforms route generation from random search to physically controlled evolution by introducing a gradient guidance mechanism based on physical power deviation within the latent space. By converting dynamic constraints into correction signals during the denoising process, the generated route helps maintain the safe linear response zone of the UAV's power actuators while avoiding terrain obstacles. This alleviates the structural contradiction of the disconnect between planned trajectory and hardware execution capabilities in traditional methods. Furthermore, by deeply coupling discrete environmental feature maps with a continuous conditional diffusion model, the route can perceive the dynamic power reserve patterns of the delivery area in real time through deviation gradients at each step of generation, thus giving rise to highly reliable adaptive routes with dynamic redundancy.

[0048] This application further proposes that, by invoking a preset saturation correction logic, and based on the deviation gradient and the power limit value, smoothing the power change rate between adjacent waypoints on the intermediate trajectory to obtain the target delivery route, the specific steps include: Saturation correction logic refers to a control algorithm that uses physical power limits to inversely constrain the trajectory evolution amplitude, specifically through saturation deviation. This is achieved through negative feedback regulation, which ensures that the generated track coordinates are dynamically feasible at the geographical level.

[0049] First, the trajectory evolution module receives the route candidate sequence. And based on the latent space deviation gradient of each track point Perform fine-tuning of the physical location. Specific steps include: Determine the correction direction: using decoding and reconstruction algorithms Mapping the latent space gradient back to the coordinate space yields the corrected direction vectors of each waypoint in the three-dimensional geographic space. This vector represents the power reserve (potential energy value) from the current coordinate to the surrounding area. The optimal gradient direction for moving a larger region.

[0050] Quantization saturation deviation: Simulating the saturation deviation at each point Move in the direction by an initial evolution step (e.g., waypoint spacing) Estimated output power after And calculate its relative to the power limit value. saturation deviation : ; in, The saturation deviation (in W) is the saturation deviation of the i-th track point. If the estimated power does not exceed the limit, the deviation is 0. To estimate output power; refers to the track point along the correction direction. Try moving an initial step size Then, the power required to maintain stable flight at the target coordinates; This refers to the power limit. It represents the maximum safe power output allowed for continuous operation of the drone's power system. This is a truncation operator. It ensures that the deviation is 0 instead of negative when the power demand is below the limit, reflecting the logic of penalizing only over-limit states.

[0051] Dynamic step size adjustment: To prevent the corrected waypoints from falling into the power over-limit region that the propulsion system cannot handle, the evolution step size is scaled according to the following formula: ; in, This is the adjusted actual evolution step size (unit: m), which is the displacement that is ultimately applied to the coordinate update; The initial evolution step size (unit: m) is the preset baseline movement increment (in this embodiment, it is usually the waypoint spacing). of ); Normalized saturation represents the proportion of power exceeding the limit to the ultimate power.

[0052] When the predicted power approaches its limit, the movement amount Approaching This ensures that the trajectory evolution always remains within the linear response range of the hardware execution capability.

[0053] After completing the coordinate correction, the system pairs adjacent waypoints. Detect and smooth the power jump slope to eliminate the risk of control failure induced by drastic changes in terrain or weather.

[0054] Slope calculation: Calculating the slope of the drone at its cruising speed During flight, the slope of the predicted power jump between adjacent points : ; in, , These are the instantaneous power requirements at waypoints i and i+1, respectively; The flight time period is obtained by dividing the Euclidean distance by the cruise speed; the estimated power jump slope is also given. It refers to a physical quantity that characterizes the rate of change of power demand between adjacent waypoints. Specifically, it is approximated by the derivative of the power difference with respect to flight time. Its function is to quantify the requirements of the flight path on the response frequency of the airborne control system.

[0055] Interval determination: judgment Is it within the attitude response bandwidth threshold? (This embodiment takes) Within the defined linear response range, which refers to the safe threshold range within which the UAV's power system can stably execute power adjustment commands, specifically determined by the response bandwidth of the flight control system. The decision serves to filter out drastic power jumps that could lead to control oscillations or failure. The upper limit of the interval is determined by the power change rate margin coefficient. (Pick )Sure: ; in, This is the power change rate margin coefficient; This is the power limit value, the safe upper limit of the power system output; The attitude response bandwidth threshold (Hz).

[0056] Spatial compensation and transition waypoint insertion: Transition waypoints are intermediate path nodes added to smooth the power curve. This is achieved through spatial interpolation and arc length compensation during the over-limit segment. Their function is to force the power variation characteristics of the trajectory back to the linear domain that the hardware can handle. If the slope exceeds the limit, the system reduces the rate of change by increasing the flight segment duration; the spatial coordinate step size is recalculated. and by inserting A number of transition waypoints are used to extend the actual flight arc length: ; in, The arc length of the spatial segment is recalculated after smoothing. The preset waypoint spacing is 30m.

[0057] The coordinates of the waypoints are calculated using linear interpolation on the horizontal plane and cubic Hermite spline interpolation on the vertical plane. Terrain gradients are incorporated into the interpolation process. As a tangent constraint, it ensures smoothness of height variations.

[0058] This application's solution constructs the final physical barrier from probabilistic flight path to airworthy flight path through deep coupling of saturation correction and power change rate smoothing. This approach not only solves the potential local jitter problem in the original sequence generated by the diffusion model, but also ensures that the final target delivery route meets the linear control requirements of the UAV's power system throughout the entire process by introducing flight control bandwidth constraints. This avoids the risk of control failure due to power over-limit or excessively rapid power jumps, transforming complex external environmental field constraints into smooth commands for the microscopic geometry of the flight path. This allows the final generated flight path to avoid high-risk areas macroscopically and adapt to the dynamic response characteristics of the underlying flight control system microscopically, thereby improving the flight safety of UAVs in extreme rural environments.

[0059] As a preferred embodiment, the solution of this application is implemented in the scenario of emergency medicine delivery in a mountainous rural area as follows: High-precision mapping drones and weather radar were used to acquire 5-meter resolution digital elevation model (DEM) data and real-time 3D wind field model data for the target delivery area (e.g., a village in a canyon terrain). The delivery mission started at a distribution center at the foot of the mountain and ended at a mountaintop medical point 5 kilometers away. Simultaneously, the dynamic parameters of the delivery drone (e.g., a hexacopter platform) were acquired: total mass... Rated power limit Total area of ​​rotor disk aerodynamic drag coefficient and attitude response bandwidth threshold .

[0060] The feature processing module was invoked to perform gradient operator operations on the DEM data, identifying terrain undulations with a maximum slope of 45° along the route; simultaneously, through flow field analysis, a canyon acceleration factor of 1.8 was extracted at the canyon narrowing point. By using spatial coordinate indexing, these terrain features are overlaid and mapped with meteorological constraints to generate environmental feature maps containing characteristics such as altitude, wind speed, and disturbance intensity. .

[0061] The potential energy modeling module combines dynamic parameters to perform physical space mapping: at altitude At this location, atmospheric density is calculated based on the international standard atmospheric model. Then calculate the hovering power. If the local wind speed achieve Then calculate the aerodynamic drag power. Approximately And superimposed disturbance compensation power According to the formula Calculate the remaining reserve power percentage of all voxels, constructing the dimension as follows: Control margin potential energy tensor The high-energy-consuming ridgeline area is visually presented as a "low-reserve zone" ( The sheltered back slope of the valley is a "high reserve area" ( ).

[0062] Will and Input conditional encoder. Extraction using 3D-CNN. 2D features And combined with the physical properties extracted by MLP Modality fusion is performed through an 8-head cross-attention mechanism to generate Multidimensional guiding condition vector At the same time, the total number of waypoints is determined based on the distance between the origin and destination. And sampled from the standard Gaussian distribution to generate Initial track noise sequence of dimension .

[0063] Perform 50 controlled denoising iterations within the latent space defined by the conditional diffusion model: Each iteration obtains the intermediate trajectory through decoding and reconstruction. And calculate instantaneous power demand. .

[0064] Some trackpoints were found attempting to cross the top of a ridge with extremely high wind speeds, resulting in... near .

[0065] Triggering deviation gradient calculation: Calculate the physical distance deviation value .

[0066] The resulting deviation gradient The predicted noise direction is corrected to guide the flight path toward the leeward slope pass where the lateral power reserve is more abundant.

[0067] After iteration, output the candidate route sequence. The trajectory evolution module invokes the saturation correction logic to calculate the saturation correction based on cruising speed. Flight, power jump slope between adjacent waypoints Compare with the upper limit of the linear response interval. The system determined that the transition in that segment was too rapid. The system automatically recalculated the spatial step size. Two transition waypoints were inserted in this segment, and cubic Hermite spline interpolation was used to smooth the altitude curve, thus restoring the power change rate to its normal value. .

[0068] The final generated target delivery route avoids physical obstacles and bypasses wind saturation zones throughout, ensuring that the drone maintains at least [a certain speed] during the complex canyon airflow. The control power margin. Compared with the traditional geometric shortest path generated by the A* algorithm, the route generated by this method, although longer, is significantly improved. However, the peak battery current decreased during task execution. This effectively avoids the risks of power drop and loss of control caused by exceeding power limits.

[0069] Through the above scheme, this application achieves dynamic adaptation of the UAV delivery route generation process to the complex physical environment of rural areas and the airborne dynamic characteristics. The introduction of a control margin potential energy tensor allows terrain and weather constraints to be uniformly quantified into perceptible power boundaries. The synergy between the conditional diffusion model and the physical deviation gradient improves the accuracy of route convergence to the safe solution space within the latent space, enhancing its adaptability to extreme physical environments compared to traditional probabilistic road network algorithms. The application of saturation correction and power smoothing improves the airworthiness of the output trajectory and reduces the pressure of high-frequency power jumps on the flight control system's response bandwidth. This adaptive mechanism improves the safety and reliability of delivery missions, making routine UAV logistics transportation possible in complex terrain and rural environments, and reducing equipment wear and accident risks under special climatic conditions.

[0070] like Figure 2 As shown, in this embodiment, an adaptive route generation system for UAV delivery in rural areas with complex terrain is provided, including the following modules: The environmental perception module is used to simultaneously acquire digital elevation model (DEM) data, 3D wind field model data, and the UAV's dynamic characteristic parameters of the delivery area. These dynamic characteristic parameters are pre-stored in the system's configuration database, including core hardware indicators such as the UAV's total mass, power limit, rotor disk area, and aerodynamic drag coefficient. The role of this module is to provide a high-fidelity digital foundation for subsequent physical modeling, ensuring that the generated flight path is based on a realistic physical environment.

[0071] The feature processing module performs spatial correlation analysis on the digital elevation model (DEM) data and the 3D wind field model data, extracting environmental feature maps containing multiple physical constraints. This module extracts elevation gradients and undulations from the DEM data using gradient operators and performs flow field analysis on the wind field data to identify key features such as canyon acceleration factors. Through spatial coordinate indexing, these heterogeneous terrain and meteorological constraints are overlaid and mapped to generate a multi-channel numerical matrix, providing the system with spatially consistent constraint inputs.

[0072] The potential energy modeling module, combined with the aforementioned dynamic characteristic parameters, maps each physical constraint term in the environmental feature map to a preset power demand space, constructing a control margin potential energy tensor characterizing the power reserve state at each spatial coordinate point within the delivery area. This module incorporates a power consumption model based on momentum theory and aerodynamic drag equations, capable of quantifying abstract environmental constraints into specific power percentage values. This mapping process enables the system to explicitly perceive the power safety margin distribution across the entire delivery area, thereby physically defining the energy plains and power no-fly zones generated by the flight path.

[0073] The condition generation module is used to perform modal fusion of the environmental feature map and the control margin potential energy tensor through a preset condition encoder, obtaining a multidimensional guiding condition vector characterizing the spatial distribution pattern of the power reserve state. This module utilizes a three-dimensional convolutional neural network to extract spatial features and calculates the correlation weights between spatial patterns and physical properties through a cross-attention mechanism. Its function is to achieve semantic alignment of heterogeneous data, transforming complex physical fields into high-dimensional semantic guiding signals recognizable by the generative model.

[0074] The data initialization module is used to obtain the preset spatial dimension of the target delivery route to be generated, and construct an initial trajectory noise sequence based on the preset spatial dimension. This module generates an initial random state with maximum uncertainty by independently and identically sampling a standard Gaussian distribution, which serves as the logical starting point for the reverse evolution of the diffusion model.

[0075] The spatial computation module is used to input the initial trajectory noise sequence and the multidimensional guidance condition vector into a preset conditional diffusion model, and to perform multi-step reverse denoising iterations within the latent space defined by the conditional diffusion model. During each denoising iteration, the module generates an intermediate trajectory and calculates the deviation gradient of the intermediate trajectory relative to the multidimensional guidance condition vector. This gradient serves as a navigation correction signal, correcting the direction of the predicted noise in real time within the latent space, guiding the trajectory to converge towards the solution space that satisfies the power constraints.

[0076] The trajectory evolution module invokes preset saturation correction logic to smooth the power change rate between adjacent waypoints on the intermediate trajectory based on the deviation gradient and the power limit value, thus obtaining the target delivery route. This module automatically compensates for the spatial step size by inserting transition waypoints by monitoring the estimated power jump slope between adjacent waypoints and comparing it with the attitude response bandwidth threshold of the flight control system. Its function is to ensure that the output final route remains within the linear control response domain of the UAV's power system throughout the entire route.

[0077] Through the above technical solutions, this application achieves end-to-end physical perception from environmental constraints to dynamic execution limits. The system maps terrain and weather limitations to a unified power demand space, alleviating the disconnect between geometrically reachable but dynamically infeasible paths in traditional path planning, enabling explicit prediction of power saturation risks in complex rural environments. At the generation logic level, controlled denoising iterations within the latent space, combined with physical deviation gradient guidance, cause the generated path to spontaneously avoid low-power reserve areas like a physical fluid, improving convergence efficiency while ensuring exploration randomness. Finally, by establishing hard constraints between the trajectory geometry and the underlying flight control bandwidth, the system can filter out power surges that induce oscillations, ensuring the flight stability of the UAV under extreme disturbances and extending the hardware lifespan of the power system.

[0078] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. An adaptive method for generating delivery routes for unmanned aerial vehicles (UAVs) in rural areas with complex terrain, characterized in that: Includes the following steps: Simultaneously acquire digital elevation model data, three-dimensional wind field model data, and dynamic characteristic parameters of the UAV for the delivery area, including preset power limit values; Spatial correlation analysis is performed on the digital elevation model data and the three-dimensional wind field model data to extract environmental feature maps containing multiple physical constraints; By combining the aforementioned dynamic characteristic parameters, each physical constraint term in the environmental feature map is mapped to a preset power demand space to construct a control margin potential energy tensor characterizing the power reserve state of each spatial coordinate point within the delivery area. By using a preset condition encoder, modal fusion is performed on the environmental feature map and the control margin potential energy tensor to obtain a multidimensional guiding condition vector characterizing the spatial distribution law of the power reserve state. Obtain the preset spatial dimension of the target delivery route to be generated, and construct an initial trajectory noise sequence based on the preset spatial dimension; The initial trajectory noise sequence and the multidimensional guiding condition vector are input into a preset conditional diffusion model. The conditional diffusion model is used to perform multi-step reverse denoising iterations in the latent space defined by the conditional diffusion model. In each denoising iteration, an intermediate trajectory is generated, and the deviation gradient of the intermediate trajectory relative to the multidimensional guiding condition vector is calculated. The preset saturation correction logic is invoked, and the power change rate between adjacent waypoints is smoothed based on the deviation gradient and the power limit value to obtain the target delivery route.

2. The adaptive route generation method for UAV delivery in complex terrain rural areas according to claim 1, characterized in that, After obtaining the target delivery route, the process also includes the following steps: Acquire real-time environmental data of the UAV during the execution of the target delivery route, and calculate the deviation of the real-time environmental data from the three-dimensional wind field model data; If the deviation exceeds a preset deviation threshold, the affected local segments in the target delivery route are identified based on the deviation, and local noise injection processing is performed on the local segments to obtain a segment to be updated containing noise features. The real-time environmental data is used as a correction and guidance condition and input into the conditional diffusion model to perform secondary diffusion and noise reduction on the flight segment to be updated, so as to update the target delivery route.

3. The adaptive route generation method for UAV delivery in complex terrain rural areas according to claim 1, characterized in that, The process of extracting an environmental feature map containing multiple physical constraints includes: Gradient analysis is performed on the digital elevation model data to extract terrain constraint terms that include altitude gradient and topographic relief; flow field analysis is performed on the three-dimensional wind field model data to extract meteorological constraint terms that include local wind speed vector and canyon acceleration factor. By using spatial coordinate indexing, the terrain constraint and the meteorological constraint are overlapped and mapped to generate an environmental feature map.

4. The adaptive route generation method for UAV delivery in complex terrain rural areas according to claim 1, characterized in that, The process of constructing the control margin potential energy tensor is as follows: A power consumption model for the UAV is constructed based on the aforementioned dynamic characteristic parameters; Using the UAV power consumption model, the estimated power consumption required for each physical constraint in the environmental feature map to maintain a stable flight attitude at the corresponding spatial coordinate points is calculated. Based on the difference between the power limit value and the estimated power consumption, the proportion of remaining reserve power at each spatial coordinate point is calculated, and a control margin potential energy tensor is generated based on the proportion of remaining reserve power.

5. The adaptive route generation method for UAV delivery in complex terrain rural areas according to claim 1, characterized in that, The process of obtaining the multidimensional guiding condition vector includes: The spatial feature representation of the environmental feature map and the physical attribute representation of the control margin potential energy tensor are extracted by the conditional encoder; the correlation weight between the spatial feature representation and the physical attribute representation is calculated by the cross-attention mechanism. Based on the association weights, the spatial feature representation and the physical attribute representation are fused to generate a multidimensional guiding condition vector.

6. The adaptive route generation method for UAV delivery in complex terrain rural areas according to claim 1, characterized in that, The process of constructing the initial track noise sequence includes: Based on the straight-line distance between the start and end points of the target delivery route and the preset waypoint spacing, the total number of waypoints N is determined; based on the total number of waypoints N, the preset spatial dimension in the original coordinate space is constructed, and the preset spatial dimension is... ; Based on the preset spatial dimension, an initial track noise sequence is generated by independently and identically sampling a standard Gaussian distribution.

7. The adaptive route generation method for UAV delivery in complex terrain rural areas according to claim 1, characterized in that, The process of performing multi-step reverse denoising iterations includes: The initial trajectory noise sequence is mapped from the original coordinate space to the latent space using a preset encoding mapping algorithm; within the latent space, controlled denoising processing based on the multidimensional guiding condition vector is performed according to a preset number of iteration steps. The latent space feature sequence after iteration is restored to the target delivery route using a preset decoding and reconstruction algorithm.

8. The method for adaptively generating delivery routes for unmanned aerial vehicles (UAVs) in complex terrain according to claim 1, characterized in that, The process of calculating the deviation gradient includes: Extract the instantaneous power demand features of each track point in the intermediate track, and establish a correlation mapping between the instantaneous power demand features and the power reserve distribution features of the corresponding coordinate points in the multidimensional guidance condition vector; The physical distance deviation of the instantaneous power demand characteristic relative to the power reserve distribution characteristic is quantified; the physical distance deviation is transformed into a gradient descent vector in the latent space to obtain the deviation gradient.

9. The method for adaptively generating delivery routes for unmanned aerial vehicles (UAVs) in complex terrain according to claim 1, characterized in that, The process of smoothing the power change rate between adjacent waypoints is as follows: Calculate the estimated power jump slope of two adjacent track points in the time domain in the intermediate state track; determine whether the estimated power jump slope is within the linear response interval limited by a preset attitude response bandwidth threshold; If it is determined that the location is outside the linear response interval, the spatial coordinate step size between the two adjacent track points is recalculated.

10. An adaptive route generation system for UAV delivery in rural areas with complex terrain, characterized by: The application includes the adaptive generation method for UAV delivery routes in complex terrain rural areas as described in any one of claims 1 to 9, comprising: An environmental perception module is used to simultaneously acquire digital elevation model data, three-dimensional wind field model data, and dynamic characteristic parameters of the UAV in the delivery area. The dynamic characteristic parameters include preset power limit values. The feature processing module is used to perform spatial correlation analysis on the digital elevation model data and the three-dimensional wind field model data, and extract environmental feature maps containing multiple physical constraint terms. The potential energy modeling module is used to combine the dynamic characteristic parameters to map each physical constraint in the environmental feature map to a preset power demand space, and construct a control margin potential energy tensor that characterizes the power reserve state of each spatial coordinate point in the delivery area. The condition generation module is used to perform modal fusion of the environmental feature map and the control margin potential energy tensor through a preset condition encoder to obtain a multidimensional guiding condition vector characterizing the distribution law of the power reserve state in space. The data initialization module is used to obtain the preset spatial dimension of the target delivery route to be generated, and to construct an initial trajectory noise sequence based on the preset spatial dimension; The spatial computing module is used to input the initial trajectory noise sequence and the multidimensional guiding condition vector into a preset conditional diffusion model, and use the conditional diffusion model to perform multi-step reverse denoising iterations in the latent space defined by the conditional diffusion model. In each denoising iteration, an intermediate trajectory is generated, and the deviation gradient of the intermediate trajectory relative to the multidimensional guiding condition vector is calculated. The trajectory evolution module is used to call the preset saturation correction logic, and according to the deviation gradient and the power limit value, to smooth the power change rate between adjacent waypoints of the intermediate trajectory to obtain the target delivery route.

Images