A method for task offloading and resource allocation in air-space-ground vehicular networks to address mobile coverage risks

By generating multi-step vehicle location predictions, calculating mobile coverage risks, decoding paths and task proportions, solving continuous resource allocation, and performing verification and rollback before execution, the system addresses mobile coverage risks and resource allocation issues in vehicles, drones, and low-Earth orbit satellite networks, thereby improving the stability and efficiency of task processing.

CN123124016APending Publication Date: 2026-11-17NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611485789.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-23
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies are insufficient to effectively address mobile coverage risks and resource constraints in a three-tiered mission processing network consisting of vehicles, drones, and low-orbit satellites, leading to mission interruptions and inefficient resource allocation.

Method used

By generating multi-step vehicle position predictions, assisting in candidate drone movements, calculating mobile coverage risks, decoding paths and task ratios, solving for continuous resource allocation, and performing verification and rollback before execution to ensure resource constraints are met.

Benefits of technology

It reduces task interruptions caused by changes in mobile coverage, ensures that resource configuration meets constraints, and improves the stability and efficiency of task processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123124016A_ABST
    Figure CN123124016A_ABST
Patent Text Reader

Abstract

This invention discloses a method for task offloading and resource allocation in a space-air-ground vehicle-to-everything (V2X) network addressing mobile coverage risk. The method predicts multi-step positions based on the vehicle's historical motion state and generates candidate UAV motions using the last-step position. The feasible domain of the link for the current cycle is determined by the current states of the vehicle, UAV, and low-Earth orbit satellite. For each candidate configuration, the mobile coverage risk is calculated based on the predicted vehicle position and the position extrapolated according to the UAV's current speed and heading. Bandwidth and computational resources are then calculated, and resource-infeasible candidates are eliminated. Before deployment, the resources of the selected configuration are recalculated based on the updated state. If the verification passes, the risk is recalculated and the entire configuration is deployed. If the verification fails, the process reverts to local processing and the remote decision and resources are cleared. The UAV and satellite branches share the vehicle-to-UAV upload mechanism. The method provided by this invention can reduce task interruptions caused by changes in mobile coverage and ensure that the deployed configuration meets resource constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of edge computing for vehicle-to-everything (V2X) networks, vehicle mobility prediction, and air-space-ground communication and computing resource allocation, specifically a method for offloading and allocating tasks and resources in air-space-ground V2X networks to address mobile coverage risks. Background Technology

[0002] Environmental perception, target recognition, and cooperative driving tasks in the Internet of Vehicles (IoV) are characterized by large amounts of input data, high computational workload, and clear completion deadlines. The computing power and available energy of the vehicle's local processor are limited. While offloading tasks to drones or low-Earth orbit satellites via wireless links can supplement computing resources, it also introduces overhead related to uploading, relaying, node computing, and resource contention.

[0003] In a three-tiered task processing network consisting of vehicles, drones, and low-Earth orbit satellites, vehicles first connect to drones, which can then relay to currently visible low-Earth orbit satellites. Since vehicles, drones, and satellites can jointly process the same decomposable task, the task processing ratio must be matched with the actual availability of the two-tier links, link bandwidth, node computing power, and the drone's remaining energy. Furthermore, continuous vehicle movement can cause currently available vehicle-drone links to approach or exceed their service range in subsequent moments. Existing solutions often use predicted locations for trajectory planning, candidate server selection, or connection interruption probability estimation, easily mixing "current feasibility" and "future stability" in the same judgment: if the path is selected solely based on the current location, it's difficult to distinguish the coverage risks caused by different processing ratios at distant points; if current candidate paths are added directly based on predicted locations, connections that haven't been established may be considered executable paths. Simultaneously, drone movement, path selection, task decomposition, and resource allocation are interdependent. Unifying discrete paths, task ratios, and continuous resources significantly increases the search dimensionality and generates resource-infeasible candidates.

[0004] Therefore, the problem that this invention urgently needs to solve is how to provide a method for task offloading and resource configuration of air-space-ground-vehicle network that takes into account both mobile coverage risks and resource constraints, calculates risks after decoding candidate paths and task ratios, solves communication and computing resources for candidate configurations, reviews the final configuration before execution, and adopts overall rollback when the review fails. Summary of the Invention

[0005] The purpose of this invention is to propose a method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risk. This method can establish a feasible link domain based on the current state, calculate mobile coverage risk based on the multi-step predicted location of the vehicle and the extrapolated location of the UAV, solve for continuous resources after risk assessment, screen feasible resource candidates, and perform resource recalculation and risk recalculation before execution to reduce task interruption caused by changes in mobile coverage and ensure that the issued configuration meets resource constraints.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] In a three-layer network consisting of vehicle terminals, drones, and low-orbit satellites, this invention generates multi-step vehicle positions based on continuous historical motion states; it assists candidate drone motions with the last-step predicted position, and obtains the current service state after boundary, spacing, and energy checks; it freezes the current link feasible domain based on the actual current vehicle position, the current service position of the drone, and the current visibility state of the low-orbit satellites; within this feasible domain, it decodes the path and task ratio, extrapolates the future position of the drone, calculates the mobile coverage risk first, then solves for continuous resources and eliminates infeasible resource configurations.

[0008] This method sequentially executes the following steps within each control cycle: "Current feasible domain freezing—candidate configuration decoding—mobile coverage risk calculation—continuous resource solving and filtering—execution-time recalculation and verification—risk recalculation and overall deployment after success or overall rollback after failure." The set of tasks processed in each cycle is evaluated according to their service status upon arrival. Vehicles do not access low-Earth orbit satellites via direct vehicle-to-satellite links. Each control cycle includes the following steps:

[0009] Step 1: Obtain the continuous historical location, speed, heading, and task attributes (i.e., current task information) of the vehicle terminal, as well as the UAV's location, speed, heading, remaining energy, and current visibility status of low-Earth orbit satellites, generating multiple vehicle location sequences for predicted time points; use the predicted location at the last time point and the current network status to generate candidate UAV motions, and obtain the current service status of the UAV after checking the service area, UAV spacing, and flight energy; establish two levels of candidate paths based on the vehicle's actual current location, the UAV's current service location, and the current visibility status of low-Earth orbit satellites, and freeze them as the current link feasible region, and freeze the current link feasible region during the candidate configuration evaluation of this control cycle; the predicted vehicle location within the feasible region is used for subsequent mobile coverage risk calculation;

[0010] Step 2: Sequentially execute candidate configuration decoding, mobile coverage risk calculation, and resource feasibility solving: Within the current link feasible domain, establish candidate configuration codes that include vehicle-UAV selection, UAV-satellite selection, and the processing ratios of local, UAV, and low-Earth orbit satellite tasks, and mask unexecutable remote processing branches based on the current link feasible domain; for each decoded candidate configuration, using the current service location of the UAV determined in Step 1 as the extrapolation starting point, obtain the future positions of the UAV at multiple predicted times according to the current speed and current heading of the UAV, and calculate the relationship between the predicted vehicle position and the future UAV position at the same predicted time. The service range out-of-bounds degree is used to calculate the mobile coverage risk, and it is weighted according to the sum of the processing ratio of UAVs and the processing ratio of low-orbit satellites. Then, the two-level path relationship and task processing ratio are fixed to solve the link bandwidth and node computing resources that meet the current link capacity, node computing power and UAV energy constraints. Candidate configurations that have not obtained resource feasible solutions are eliminated. Only for candidate configurations that have obtained resource feasible solutions, the resource configuration cost and the mobile coverage risk are combined for ranking. The vehicle predicted location and mobile coverage risk do not add, replace or restore the path in the current link feasible domain, and do not allow resource infeasible candidates to enter the ranking.

[0011] Step 3: Before the configuration is distributed, the selected configuration is recalculated and verified based on the updated actual vehicle location, actual UAV service status, current visibility status of low-orbit satellites, and remaining node resources. If all verifications pass, the final mobile coverage risk is recalculated based on the updated predicted vehicle location and the future location extrapolated from the actual UAV service location, current speed, and current heading. Then, the two-level association, three-branch task ratio, link bandwidth, and node computing resources are distributed as a whole. If any verification fails, no remote configuration is retained. Instead, the task is rolled back to all local processing, and the two-level association, remote task ratio, and all remote resources are cleared simultaneously, while the flight energy consumption that has already occurred is retained.

[0012] Preferably, the predicted position at the last moment in the vehicle position sequence in step one adopts a residual sequence to sequence gated cyclic unit model, which takes the vehicle's position, speed and heading at consecutive historical moments as input, and superimposes the predicted residuals based on the constant speed extrapolation position; the predicted last step position participates in the generation of candidate motions for the UAV, and the complete predicted position sequence at the last moment is used for subsequent calculation of mobile coverage risk for candidate configuration.

[0013] Preferably, the UAV trajectory control strategy in step one generates acceleration and heading increments based on its own state, the vehicle's current relative position and predicted final relative position, the vehicle's mission, the states of other UAVs, and the visibility status of low-orbit satellites. The current service state of the UAV is determined through bounded motion mapping, velocity clipping, service area clipping, UAV spacing checks, and flight energy checks. For mobile coverage risk assessment, the future position sequence of the UAV is extrapolated from this current service location using the checked current velocity and heading, and this sequence is used for risk calculation of candidate configurations within the current link feasible domain.

[0014] Preferably, for K vehicles and M drones, the candidate configuration encoding includes 3K+M normalized components, which only represent the two-level path selection and three-branch task splitting within the current feasible domain of the link; the link bandwidth and node computing resources are not included in the encoding, but are solved separately after the path relationship and task processing ratio of each candidate configuration are fixed. Among them, K components select the drones associated with the vehicles, M components select the low-orbit satellites associated with the drones, and the other 2K components determine the logarithmic weights of the task splitting for vehicle local processing and drone processing, while the logarithmic weight of the task splitting for low-orbit satellite processing is fixed to zero.

[0015] Preferably, in step two, the mobility coverage risk is first calculated for the decoded candidate configurations. Then, under the condition of fixed two-level path relationships and task processing ratios, the resource allocation problem is solved with link bandwidth and node computing resources as continuous variables. Only when a finite solution is obtained that satisfies the constraints of positive bandwidth of active links, positive computing resources of active processing branches, total node capacity, and remaining energy of UAVs, is the candidate configuration added to the resource feasible candidate set, and the resource allocation cost is merged with the calculated mobility coverage risk to form a fitness. Candidate configurations not added to the resource feasible candidate set are not accepted and are not re-entered into the sorting due to a small mobility coverage risk value.

[0016] Preferably, in step two, starting from the current service location of UAV m determined in step one, and according to the current speed and heading of the UAV, constant speed and constant heading are extrapolated at control cycle intervals to obtain the future position of UAV m in the h-th prediction step; using the three-dimensional distance between the position of vehicle k in the h-th prediction step and the future position of the UAV in the same prediction step, the maximum service distance of the UAV, and the number of prediction steps H, the link prediction coverage risk is calculated:

[0017] ;

[0018] in, Let represent the three-dimensional distance between the predicted position of vehicle k at the h-th prediction time and the extrapolated future position of drone m. This indicates the fixed coverage radius preset by the drone m. =max{x,0}, where, for .

[0019] Associating vehicles with drones and remote processing ratio = + Weighted prediction of candidate configuration coverage risk:

[0020] ;

[0021] in, This variable indicates the association between vehicle k and drone m; it takes the value 1 when the two are associated, and 0 otherwise. and These represent the proportions of processing by drones and satellites, respectively; all tasks are executed locally. =0; For candidate configurations that satisfy resource and energy constraints, G+ As the fitness, G is the weighted cost obtained from convex optimization. Assign positive infinite fitness to candidate resource-infeasible configurations with a predicted risk weight greater than zero.

[0022] Preferably, in step two, the resource allocation cost is obtained by weighting the task completion delay, system energy consumption, and deadline timeout, where the deadline timeout is the positive part of the difference between the task completion delay and the task deadline; the task completion delay is the maximum value of the delays of each active processing branch among the local processing branch, the UAV processing branch, and the low-Earth orbit satellite processing branch; the UAV processing branch and the low-Earth orbit satellite processing branch share the same vehicle-UAV upload, and the uploaded data volume is the product of the sum of the processing ratios of the two remote tasks and the task input data volume; the UAV aggregates and forwards the satellite data to be processed associated with the vehicle, and the delay of the low-Earth orbit satellite processing branch includes the shared upload delay, the aggregation and forwarding delay of the UAVs, the UAV-satellite propagation delay, and the satellite computing delay; the system energy consumption is the sum of the vehicle local computing energy consumption, vehicle-UAV upload energy consumption, UAV computing energy consumption, UAV forwarding energy consumption to low-Earth orbit satellite, UAV flight energy consumption, and low-Earth orbit satellite computing energy consumption generated during the completion of the vehicle task in the current control cycle, wherein the shared upload energy consumption and aggregation and forwarding energy consumption of the vehicle-UAV are each counted only once.

[0023] Preferably, in step two, the continuous resource allocation problem uses the bandwidth of each active link, the computing resources of each active remote processing branch, the task completion delay auxiliary variable, and the deadline timeout auxiliary variable as optimization variables. The task completion delay auxiliary variable is set to be no less than the delay of each active processing branch of the vehicle, and the deadline timeout auxiliary variable is set to be no less than the difference between the task completion delay and the task deadline and no less than zero. The continuous resource allocation problem satisfies the bandwidth and computing resource capacity constraints of each UAV and low-orbit satellite, the positive resource conditions of the active branches, and the constraint that the sum of the UAV's flight, computing, and forwarding energy consumption does not exceed the remaining energy. The resource allocation corresponding to inactive links and processing branches is zero.

[0024] Preferably, in step two, during the initialization of the candidate configuration search, all locally processed reserved configurations are set. The updated candidate configurations are always re-decoded based on the frozen current link feasible domain, and the mobile coverage risk calculation and continuous resource solution are executed sequentially. In step three, the candidate configurations selected from the sorting results in step two are used as the configurations to be issued. The configurations to be issued are only recalculated based on the updated vehicle predicted position and the future position of the UAV extrapolated according to the current speed and current heading of the UAV after all resource recalculation and strict numerical tolerance verification before the configuration is issued. If any verification fails, the vehicle local processing ratio is set to one, the UAV and low-orbit satellite processing ratio is set to zero, and the two-level association and all remote resources are set to zero. The UAV flight energy consumption already included in the energy budget remains unchanged.

[0025] Preferably, availability tags are set for local, UAV, and LEO satellite processing branches based on the frozen current link feasible domain, and the processing ratio of the three branches is obtained through proportional mapping with availability tags. For each decoding configuration, the mobile coverage risk is first calculated based on the positions of the vehicle and UAV at each prediction step, and then the continuous resources are solved. Only when the resource solution results satisfy the constraints of positive active resource value, node capacity, and UAV energy are the resource configuration cost and the risk combined for candidate ranking. The risk is used for ranking among resource feasible candidates, and the current coverage and current visibility conditions are determined by the current link feasible domain.

[0026] Preferably, the local processing branch is executed independently, while the UAV processing branch and the low-orbit satellite processing branch share the same vehicle-UAV upload. The UAV aggregates the satellite data to be processed from its associated vehicle and then forwards it. The task completion time is the maximum of the completion times of each active processing branch. The vehicle-UAV upload latency and upload energy consumption are calculated once based on the total data volume of the two remote branches, and the UAV aggregation and forwarding energy consumption is also included once.

[0027] Preferably, after fixing the two-level path relationship and task processing ratio, a continuous resource allocation problem is constructed using task completion delay and deadline timeout auxiliary variables. The link bandwidth and computing resources are determined under the constraints of bandwidth, node computing power and UAV remaining energy. The candidate configuration search retains all local processing configurations. Before the selected configuration is executed, the resources are recalculated and verified with the updated actual state. If it passes, the final mobile coverage risk is recalculated and distributed as a whole. If it fails, the remote decision and resources are cleared as a whole.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] (1) The current link feasible domain is generated from the actual current state and remains unchanged during the candidate evaluation period of this cycle; the predicted locations of vehicles and drones are used for the mobile coverage risk assessment of candidate configurations within the feasible domain.

[0030] (2) Covering risks by considering both the multi-step predicted position of the vehicle and the multi-step future position of the drone based on the current speed and current heading, is more in line with the air-space-ground vehicle network scenario where both the vehicle and the service drone are in motion.

[0031] (3) The coverage risk of candidate configurations is calculated first, and then the continuous resources are solved; candidates that meet the bandwidth, computation and energy constraints are sorted.

[0032] (4) The UAV and satellite processing branches share a single vehicle-UAV upload; the selected scheme is recalculated and verified before being issued, and the risk is recalculated and issued as a whole after it passes. If it fails, all will be rolled back to local, remote decision-making and remote resources are cleared, while retaining the UAV flight energy consumption that has already occurred. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the three-layer task processing architecture of vehicle, drone and low-orbit satellite described in this invention;

[0034] Figure 2 This is the overall flowchart of the current link feasible domain freezing, mobile coverage risk calculation, resource solution screening, execution review and overall rollback described in this invention;

[0035] Figure 3 This is a schematic diagram illustrating the three-branch processing, shared upload, and low-orbit satellite relay relationships of the vehicle mission described in this invention. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment employs discrete control cycles for motion decision-making and resource allocation. The completion delay, computation, and communication energy consumption of each task are estimated based on their arrival cycle correlation and resource scheme, belonging to the decision evaluation model of the current task set. The control cycle is not equivalent to the task completion deadline; the expected completion delay of a task can exceed one control cycle. In the following formulas, variables without explicitly labeled time refer to the current service status to be evaluated.

[0037] Example 1

[0038] In this embodiment, the arriving task set is processed according to the control cycle, and resource constraints are established for the tasks in the current cycle; energy recursion is used to update the available energy in each cycle, and subsequent cycles are solved again based on the newly arriving tasks and the updated status.

[0039] The network in this embodiment is as follows: Figure 1 As shown, the system includes K vehicles, M drones, and S low-Earth orbit (LEO) satellites. Each vehicle has a local processor, the drones have access to communication, computing, and satellite relay capabilities, and the LEO satellites have computing capabilities. Each vehicle can be associated with at most one drone within a control cycle, and each drone can be associated with at most one currently visible satellite. One drone can serve multiple vehicles, and one satellite can provide resources to multiple drones. Vehicles do not offload tasks via direct vehicle-satellite links.

[0040] The controller acquires vehicle position, speed, heading, and task input data. Calculation of workload and task deadline Obtain the drone's location, speed, heading, and remaining battery power. The system includes satellite position, communication, and computing capacity. Vehicle position and motion status are provided by vehicle terminal status information, while node resources and satellite visibility are provided by the corresponding node status information. The controller can be handled by a computing node that maintains information interaction with the UAV, and an independent policy network can be deployed on the corresponding UAV; this deployment arrangement does not change the following hierarchical solution process.

[0041] Vehicle tasks are tasks that can be broken down according to the proportion of input data and computational workload, with local, drone, and satellite processing proportions as follows: , and All three are non-negative and sum to one. The computational workload is expressed in the number of CPU cycles required. In the cost calculation of this embodiment, the amount of task result data is relatively small compared to the amount of input data, and the latency and energy consumption of result return are ignored.

[0042] like Figure 2As shown, this embodiment includes steps 201 to 203; step 201 generates and freezes the current link feasible domain; step 202 completes candidate configuration decoding, mobile coverage risk calculation and continuous resource solution filtering in sequence within the feasible domain; step 203 completes resource recalculation verification during execution, risk recalculation and overall distribution after passing, or all local rollback after verification failure.

[0043] Step 201: Predict vehicle location, determine the current service status of the drone, and freeze the current feasible domain of the link:

[0044] Step 201-1: Vehicle multi-step position prediction: Establish a length of [missing information] for each vehicle. The historical state window contains four features: two-dimensional position, velocity, and heading. Position is normalized according to the length of the service area, velocity is normalized according to the training velocity scale, and heading is normalized according to π. The samples are divided into training, validation, and testing parts in chronological order.

[0045] A sequence-to-sequence gated cyclic unit model is constructed using a two-layer encoder and a two-layer decoder, with 128 hidden units in each layer. The encoder reads the history window and passes the end hidden state to the decoder. The decoder takes the current position as input in its first step and the predicted position from the previous step as input in subsequent steps, outputting a two-dimensional residual for the corresponding constant-velocity extrapolation position, resulting in:

[0046] ;

[0047] in To predict the location, For constant speed extrapolation position, Let be the residual output by the decoder, and t be the prediction start time. This indicates that the position at time t is predicted for the h-th future prediction time, where H is the number of prediction steps. In this embodiment, the increments of the two most recent historical positions are used as the reference displacement for each prediction step, and the current position is multiplied by h to form a constant-velocity extrapolation position. The prediction model is trained using the mean square error between the predicted and actual positions over multiple steps, and is stopped early based on the validation error. In this embodiment, we take... =10, H=5.

[0048] The prediction model outputs results in normalized coordinates, which are multiplied by the service area length to restore physical coordinates in meters. After cropping to the service area, the relative position and 3D link distance are calculated. The final predicted position is incorporated into UAV policy observations to generate candidate motions. The complete H-step vehicle prediction sequence is saved separately and used in step 202, along with the UAV future position sequence, for calculating the mobility coverage risk of candidate configurations. After offline training, the prediction model performs forward inference in each control cycle.

[0049] Step 201-2: Generation and inspection of candidate motions for UAVs: observation of the UAV m-structure. This includes the vehicle's own position, speed, heading, and normalized remaining energy; the current relative position, speed, heading, mission attributes, and final predicted relative position of each vehicle; the relative status of other UAVs; and the visibility of each satellite. Vehicle information is sorted by a fixed index, with zeros padded for any portion not exceeding the maximum number of vehicles configured.

[0050] Each UAV is configured with an independent policy network, while a shared centralized value network is used during training. Each policy network consists of three fully connected hidden layers, each containing 256 neurons and employing the ReLU activation function. It then outputs the shape parameters of the Beta distribution for the two action dimensions via α and β parameter heads, respectively. The shared value network takes into account the global state formed by stitching together observations from each UAV, as well as aggregated task load and satellite availability information. It also consists of three fully connected hidden layers, each containing 256 neurons and employing the ReLU activation function, and outputs the global state value. The shape parameters of each Beta distribution are determined by "1 + softplus(network output)", where softplus(x) = ln(1 + exp(x)). The normalized actions are mapped as follows:

[0051] ;

[0052] in, This indicates the upper limit of the acceleration amplitude preset for the drone m, which is a parameter preset based on the drone's dynamic capabilities. This represents the pre-set upper limit of the heading change amplitude within a single control cycle; it is the pre-set maximum heading increment for a single cycle. , Located in [0,1], samples are taken from the Beta distribution during training; the mean of the corresponding Beta distribution is used during execution. The UAV's speed, heading, and 2D position are updated according to the control cycle length Δt:

[0053] ;

[0054] ;

[0055] Here, clip represents range clipping. and These represent the flight speed and heading angle of the UAV before the update, respectively. This represents the acceleration obtained from the action mapping of the policy network. This indicates the maximum flight speed preset for the drone. Indicates the length of the control cycle; Indicates to Perform interval The candidate flight speeds obtained after cropping; Indicates the increase in course Superimposed on the current heading angle The candidate heading angles were then obtained and adjusted accordingly. Take the mold; and Let m represent the two-dimensional horizontal position vectors of the UAV before and after the update, respectively; and These represent the directional components of the candidate heading in the two horizontal coordinate directions, respectively, with superscripts indicating the directional components. This represents the vector transpose. The two-dimensional position variable represents the UAV's horizontal position, the flight altitude is kept at a preset value, and the heading variable is modulo 2π. The above motion update first forms candidate service states, then prunes candidate coordinates that are outside the service area, and checks the UAV's minimum spacing and flight energy budget; only candidate service states that pass the check are fixed as the UAV's current service state for this control cycle, and their positions are used to establish the current link feasible region. When calculating the mobile coverage risk, the current service position is used as the extrapolation starting point, keeping the current speed and current heading unchanged after the check, and the UAV's future position is extrapolated step by step according to the prediction step size; therefore, the UAV is not stationary in the prediction time domain. This embodiment uses the following command motion energy consumption approximation formula:

[0056] ;

[0057] in For basic propulsion and airborne equipment power greater than zero, and This is the effective maneuvering energy consumption coefficient. The flight energy consumption for the current cycle is first included in the budget, and the lower layer only allocates computing and forwarding resources from the remaining budget. This embodiment ensures that each participating UAV meets the following requirements. ≤ The solution is obtained under the given conditions; if the flight itself does not meet these conditions, it does not belong to the service state that can be restored by local unloading and rollback in this embodiment. This article does not equate zero computation and zero forwarding energy consumption with zero propulsion power consumption.

[0058] Step 201-3: Establishment and Freezing of the Current Link Feasibility Region. For the currently fixed service location of the UAV after inspection, the vehicle-UAV current path condition is defined as follows: the three-dimensional distance between the vehicle's actual current location and the UAV does not exceed the maximum service distance; the UAV-satellite current path condition is defined as follows: the current elevation angle between the UAV and the satellite is not less than the minimum elevation angle. Here, is the corresponding horizontal distance, and atan2 is the two-parameter arctangent. When the horizontal distance is zero and the satellite is higher than the UAV, the elevation angle is π / 2. The two-level paths satisfying the above conditions constitute the current link feasibility region, and ordered candidate sets are formed respectively.

[0059] The current feasible region of the link is generated from the vehicle's actual current location, the UAV's current service location after inspection, and the current visibility status of the low-Earth orbit (LEO) satellites, and remains unchanged during candidate decoding, risk calculation, and resource solving within the current control cycle. The predicted vehicle location and the extrapolated UAV location are used to calculate the coverage risk of candidate paths within the feasible region. The LEO satellite locations can be provided by a given trajectory sequence; in the simulation, equivalent periodic projection trajectories at a fixed altitude can be used to calculate the link distance and elevation angle.

[0060] Step 202: Decode candidate configurations within the frozen current link feasible domain, first calculate mobile coverage risk, then solve for continuous resources and filter candidates.

[0061] Step 202-1: Candidate Configuration Decoding. Receive the current service status of the UAV, the complete vehicle prediction sequence, and the frozen current link feasible region determined in Step 201. For each candidate configuration, the execution order is fixed as follows: decode the two-level path and task ratio; extrapolate the UAV's future position sequence from its current service location based on its current speed and heading; calculate the mobile coverage risk; subsequently solve for continuous resources and determine the current resource feasibility; only for resource-feasible configurations, combine resource costs and risks to calculate fitness. Configurations that fail to solve for resources are immediately removed, and their risk values ​​do not affect the removal result. Each candidate configuration is represented as:

[0062] ;

[0063] Where i represents the index of the candidate configuration. This represents the encoding vector of the i-th candidate configuration; This represents the vehicle-drone association selection subvector, with one component corresponding to each vehicle; This represents the selection sub-vector for the association between UAVs and low-Earth orbit satellites, with one component corresponding to each UAV; This represents the task split encoding sub-vector representing the local processing branch of the vehicle; This represents the task breakdown encoding sub-vector for the UAV processing branch. The task breakdown logarithmic weight for the low-Earth orbit satellite processing branch is fixed at zero, therefore no separate encoding sub-vector is set for it. It has K components, It has M components, Each has K components. For a set of n candidate associations in the current feasible domain of the link, the normalized association component x is mapped to the corresponding index; no association is established when the set is empty. This yields the binary vehicle-UAV association and the UAV-satellite association. Even if a predicted location enters the service range of other nodes during this period, the index set remains unchanged.

[0064] Convert the split components to = -8 + 16 , = -8 + 16 , =0, where and Let represent the local processing coding component of vehicle k and the UAV processing coding component in candidate configuration i, respectively, both of which are located in the interval [0,1]. , and Let represent the logarithmic weights of the local processing branch, the UAV processing branch, and the low-Earth orbit satellite processing branch for vehicle k, respectively. This indicates the availability of branch j, where j∈{L,U,S} represents a task processing branch. These represent the local processing branch, the UAV processing branch, and the low-Earth orbit satellite processing branch, respectively. Local branch =1, when there are no vehicle candidate drones = =0; When the vehicle has selected a candidate drone, but the selected drone has no visible candidate satellites. =, where This indicates whether the processing branch j for vehicle k is available; it is set to 1 if available and 0 otherwise. The proportion is formed by the following formula:

[0065] ;

[0066] in, This represents the initial task proportion assigned to processing branch j for vehicle k before thresholding and renormalization; r represents the summation index of the processing branches in the denominator, which iterates through the three processing branches: local, UAV, and low-Earth orbit satellite. This represents an exponential function. The threshold value will be less than a preset threshold. Set the proportion to zero, then normalize the remaining proportions; if all proportions are set to zero, retain the available branch with the largest proportion before zeroing. Clear vehicle-drone associations when there is no remote processing load; clear satellite associations when a drone has no satellite data to process. Do not allocate offload resources to inactive vehicles.

[0067] After decoding, the following conditions are met: each vehicle can be associated with at most one drone, each drone can be associated with at most one low-Earth orbit satellite, and the sum of the processing ratios for local, drone, and low-Earth orbit satellite tasks is one, with each processing ratio not less than zero. All local execution schemes are reserved separately for initialization, and they output one precisely under finite weights without relying on softmax. At this point, the discrete paths and task ratios have been determined, and the mobile coverage risk weighted by the remote processing ratio can be calculated accordingly. Link bandwidth and computing resources are then solved separately.

[0068] Step 202-2: Mobility Coverage Risk Calculation. For candidate configuration i, each step in the complete prediction sequence adopts the same mobility state at the same prediction time: the vehicle position is taken as the predicted vehicle position; the UAV position is taken as the current service position determined in step 201, and is extrapolated outward at a constant speed and constant heading according to the checked current speed and current heading, with the prediction step size remaining constant. Its horizontal position is equal to the sum of the current service position and the displacement along the current heading within the prediction time, while the altitude remains unchanged. The three-dimensional prediction distance between the vehicle and the UAV is defined as follows:

[0069] ;

[0070] ;

[0071] When the denominator is zero, the predicted coverage risk of the candidate configuration is set to zero. Let G be the resource configuration cost obtained after successful resource solution and verification. The weighted sum of the resource configuration cost and the predicted coverage risk is used as the fitness of the candidate configuration, with the predicted risk weight being greater than zero. Configurations that fail the current link feasible domain decoding or continuous resource feasibility screening are directly assigned positive infinite fitness and are not restored due to a smaller risk term. The risk is the cost term exceeding the vehicle-UAV service distance in the future; it does not represent the predicted interruption probability, does not participate in the current feasible domain membership determination, and does not replace the current visibility determination of low-Earth orbit satellites.

[0072] Step 202-3: Continuous Resource Allocation and Feasibility Screening. After fixing the association and task ratio, calculate the link transmission rate. In this embodiment, vehicle-UAV and UAV-satellite use different spectrums, parallel links at the same layer are allocated orthogonal bandwidths, a large-scale channel gain g is used, and inter-link interference is ignored. The transmit power p, bandwidth b, and noise power spectral density are... ,use:

[0073] ;

[0074] Where R(b) represents the achievable transmission rate of the link when the allocated bandwidth is b; b represents the bandwidth allocated to the link; p represents the transmit power of the link; and g represents the large-scale channel power gain of the link. Represents the noise power spectral density. This represents the noise power within bandwidth b.

[0075] Inactive links have their bandwidth and speed set to zero. Vehicle-to-drone gain can be achieved using... ,in, The large-scale channel power gain of the link from vehicle k to UAV m is represented; k and m represent the vehicle index and the UAV index, respectively. This represents the reference channel power gain of the vehicle-to-unmanned aerial vehicle (V2V) link at a unit reference distance. This represents the three-dimensional distance between the current service locations of vehicle k and drone m; Indicates the path loss index of the vehicle-to-drone link; superscript This indicates the vehicle-to-drone link type. It is evident that the drone-to-satellite link uses... ,in, This represents the large-scale channel power gain of the link from the UAV m to the low-Earth orbit satellite s; m and s represent the UAV index and the low-Earth orbit satellite index, respectively. This represents the reference channel power gain of the UAV-Low Orbit satellite link at a unit reference distance. This represents the three-dimensional distance between the drone m and the low-orbit satellite s; Indicates the path loss index of the UAV-Low Orbit satellite link; superscript This indicates the type of UAV-Low Orbit satellite link. ν and ν represent the reference gain and path loss exponent for the corresponding link, respectively. In the inner-layer solution, the transmit power, link geometry, and mission ratio are given quantities.

[0076] like Figure 3 As shown, for the task of vehicle k, with For local CPU frequency, For the selected vehicle-drone upload rate, in and For the drone and satellite computing resources allocated to this vehicle, we obtain:

[0077] ;

[0078] ;

[0079] in Indicates the local processing branch latency of vehicle k; , and These represent the proportions of vehicle k's tasks processed locally, by drone, and by low-Earth orbit satellite, respectively. This represents the computational workload required to complete all tasks for vehicle k, i.e., the number of CPU cycles. Indicates the local CPU frequency of vehicle k; This represents the upload latency of the vehicle-UAV shared by the two remote processing branches of vehicle k. This represents the amount of task input data for vehicle k; This represents the upload rate from vehicle k to its associated drone. Since the drone and low-Earth orbit satellite processing branches share the same vehicle-drone upload, the upload data volume is... . This represents the total latency of vehicle k's task on the UAV processing branch; This indicates that the vehicle and the drone share the same upload latency; This indicates the number of CPU cycles required for the portion of the task assigned to the drone for processing. This represents the computing resources, i.e., CPU frequency, allocated to vehicle k by the associated drone. Conditions This indicates that the drone processing branch is active; when When that branch's latency is zero, it is not included in the calculation of the maximum task completion latency. When both remote branches are inactive, =0; When a processing branch is not active, its branch latency is set to zero and is not included in the calculation of the maximum completion latency. The UAV m aggregates the satellite data to be processed from its associated vehicles into batches.

[0080] ;

[0081] in This represents the amount of aggregated data that the drone m collects from its associated vehicle and is ready to forward to a low-Earth orbit satellite; This variable indicates the association between vehicle k and drone m; it takes the value 1 when the two are associated, and 0 otherwise. This represents the amount of data that vehicle k is allocated to the low-Earth orbit satellite for processing. This indicates the time delay required for the drone m to forward the aggregated data to its associated low-Earth orbit satellite; This represents the transmission rate from the drone (m) to its associated low-Earth orbit satellite. When season The fluid forwarding approximation within the control cycle is adopted, and the aggregation service duration is used to evaluate each relevant satellite branch; this approximation does not additionally introduce the unified batch start time after all vehicles have completed uploading. Let vehicle k connect to satellite s via UAV m, then:

[0082] ;

[0083] ;

[0084] in This represents the total latency of the low-orbit satellite processing branch for vehicle k. This represents the shared upload latency from vehicle k to its associated drone; This indicates the latency for the associated drone m to forward the aggregated data to the low-Earth orbit satellite; This represents the three-dimensional distance between the drone m and its associated low-orbit satellite s; Represents the speed of light. This indicates the propagation delay of the UAV-Low Orbit satellite link; This indicates the proportion of low-orbit satellite processing for vehicle k. This represents the number of CPU cycles required for all tasks of vehicle k. This represents the computing resources allocated by the low-Earth orbit satellite s to vehicle k. (Condition) This indicates that the low-orbit satellite processing branch is active. This indicates the task completion delay for vehicle k; j represents the task processing branch index. These represent the local processing branch, the UAV processing branch, and the low-Earth orbit satellite processing branch, respectively. This represents the branch delay of vehicle k on processing branch j; This indicates the proportion of tasks assigned to processing branch j; This represents the set of active processing branches for vehicle k, containing only branches with a task processing ratio greater than 0. Since each active processing branch executes in parallel, the vehicle task completion delay is the maximum value of the delays of all active processing branches. The vehicle-to-UAV propagation delay is relatively small compared to the above items and is ignored in this embodiment. The above task completion delay expression reflects that the local branch executes independently, and the two remote branches perform downstream processing separately after being jointly uploaded.

[0085] by , and These represent the effective energy consumption coefficients of the vehicle, drone, and satellite processor, respectively, for calculating local energy consumption. Upload power consumption ,in This represents the locally calculated energy consumption of vehicle k; This represents the effective energy consumption coefficient of the vehicle's local processor k. This represents the number of CPU cycles allocated to the local processing branch for vehicle k; Indicates the local CPU frequency of vehicle k; This represents the upload energy consumption of vehicle k; This represents the upload transmit power of vehicle k; This indicates that the vehicle and the drone share the same upload latency. And:

[0086] ;

[0087] ;

[0088] ;

[0089] in This represents the computational energy consumption of the drone m in processing its associated vehicle tasks. The variable indicating the association between vehicle k and drone m; This represents the effective energy consumption coefficient of the drone's m-processor; This represents the number of CPU cycles allocated to the UAV processing branch for vehicle k; This represents the computing resources allocated by drone m to vehicle k; This represents the energy consumption for data forwarding by drone m. This indicates the transmission power of the drone m when sending data to a low-Earth orbit satellite. This represents the latency of the aggregated data forwarding for drone m. This represents the computational energy consumption of the low-orbit satellite s in processing vehicle tasks relayed by various drones; The variable indicating the association between the drone m and the low-orbit satellite s; This represents the effective energy consumption coefficient of the processor for low-Earth orbit satellites. This represents the number of CPU cycles allocated to the low-Earth orbit satellite processing branch for vehicle k. This represents the computing resources allocated by low-orbit satellite s to vehicle k; product Only tasks transmitted via the two-stage path of vehicle k—UAV m—Low-Earth Orbit satellite s are counted. The system energy consumption E is obtained by summing the local computing, uploading, and satellite computing energy consumption of all vehicles with the computing, forwarding, and flight energy consumption of all UAVs. Uploading energy consumption is counted once per vehicle, and forwarding energy consumption is counted once per UAV.

[0090] Let the current active vehicle group be , This represents the number of vehicles used to calculate the average value, where Represents a set The number of vehicles in the vehicle. Let... For fixed delay normalized reference value, To establish a normalized reference value for system energy consumption, construct the current cycle cost:

[0091] ;

[0092] ;

[0093] Where C represents the weighted cost of the current control cycle; This represents the set of active vehicles with pending tasks within the current control cycle; Indicates the task completion delay for vehicle k; This represents the preset time delay normalized reference value; E represents the total system energy consumption in the current control cycle; This represents a pre-set normalized reference value for system energy consumption; This represents the trade-off coefficient between normalized task latency and normalized system energy consumption; This indicates the amount of violation of the normalization cutoff deadline; This represents the penalty coefficient for violations within the deadline. This represents the average normalized deadline violation rate for active vehicle tasks within the current control period; Indicates the deadline for vehicle k's mission; This indicates the duration by which the task completion time exceeds the task deadline; Denotes the positive part function, for .when When vehicle k contributes 0 to the amount of violation by the deadline; when In this case, the contribution is the ratio of timeout duration to task deadline. This embodiment uses the timeout amount as a soft penalty; all local processing can still be evaluated, but it does not guarantee that all tasks will meet their deadlines. The long-term optimization objective is the average cost of continuous control cycles.

[0094] To solve for the resource allocation of a given particle, we introduce... ≥ The constraint scope covers all activity branches; at the same time, it introduces... ≥ - , ≥0, and with and The alternative cost formula includes task latency and positive timeout. Solving variables include vehicle-UAV bandwidth. Unmanned Aerial Vehicles (UAVs) - Satellite Bandwidth UAV computing resources Satellite computing resources and auxiliary variables , .

[0095] Resource allocation satisfies:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] in This represents the bandwidth allocated to the link from vehicle k to drone m; This represents the vehicle-drone link bandwidth capacity managed by drone m. The variable indicating the association between vehicle k and drone m; This represents the bandwidth allocated to the link between the UAV m and the low-Earth orbit satellite s; This indicates the bandwidth capacity of the UAV-satellite link managed by the low-Earth orbit satellite s; The variable representing the association between the UAV m and the low-Earth orbit satellite s. The above constraints limit the total bandwidth allocation for the access link and the satellite link, respectively, and fix the bandwidth of links without established association to 0. This represents the computing resources allocated to the drone processing branch of vehicle k for drone m. This represents the computational resource capacity of drone m; This represents the computing resources allocated by low-Earth orbit satellite s to the low-Earth orbit satellite processing branch of vehicle k; This represents the computational resource capacity of the low-Earth orbit satellite *s*. (Product) This indicates whether a two-level path exists for vehicle k to access low-Earth orbit satellite s via drone m; when no usable two-level path exists, the corresponding low-Earth orbit satellite computing resources are fixed at 0. This represents the flight energy consumption of the drone m during the current control cycle; This represents the computational energy consumption of the drone m in processing its associated vehicle tasks. This refers to the energy consumption generated by the drone m forwarding satellite data to a low-Earth orbit satellite. This represents the remaining energy budget for UAV m at the start of the current control cycle. This constraint ensures that the sum of the UAV's energy consumption for flight, computation, and relay within the current control cycle does not exceed its remaining energy. When the task involves remote processing, the upload bandwidth must be positive; when the UAV's processing volume is positive, its CPU resources must be positive; when the satellite's processing volume is positive, both relay bandwidth and satellite CPU resources must be positive. Inactive resources are fixed at zero.

[0102] For fixed associations, fixed task ratios, and given link geometry, the above link transmission rate expression is a positive concave function in the positive bandwidth domain, and its reciprocal constitutes a convex delay term. The computational delay and dynamic computational energy consumption are also convex functions. Since the aggregated forwarding data volume is fixed, the forwarding delay and energy consumption are also convex functions of the bandwidth. Combining bandwidth, computational resources, energy, and auxiliary variable constraints, a continuous resource allocation problem is formed. This problem is solved after calculating the mobile coverage risk of candidate configurations; after solving, the numerical finiteness of the solution, the positive values ​​of all active resources, and the constraints of total node resources and energy are verified. Configurations that fail verification are directly marked as infeasible and not included in the sorting; only for configurations that pass verification, the resource allocation cost is merged with the previously calculated mobile coverage risk to form a fitness.

[0103] Step 202-4: Candidate Configuration Update and Sorting. Based on the individual's historical best position. and the global historical best position Update particles:

[0104] ;

[0105] ;

[0106] ;

[0107] in and Let represent the position vector and velocity vector of particle i at the t-th iteration, respectively; the particle position vector is the candidate configuration code. This represents the best historical position of particle i up to the current iteration. This represents the global historical best position obtained by all particles up to the current iteration; This represents the inertia weight used in the t-th iteration; and These represent the individual learning coefficient and the global learning coefficient, respectively. and It represents a random vector with the same dimension as the particle position vector, whose components are independent and follow a uniform distribution in the interval [0,1]. This indicates that the corresponding components are multiplied element by element; This means clipping each component of the vector to the interval [0,1]. for After each update, the system is re-decoded based on the same frozen current link feasible domain. First, mobile coverage risk is calculated, then continuous resources are solved. After resource solving and verification pass, the resource configuration cost and risk are combined to calculate fitness. Risk weights, branch thresholds, number of candidates, and number of iterations are pre-set as deployment configurations and remain consistent throughout the same evaluation round.

[0108] Step 203: Recalculate and verify resources based on the actual status updated during execution. If successful, recalculate risks and distribute the entire resource. If unsuccessful, roll back to local processing.

[0109] Step 203-1: Execute status refresh and resource recalculation verification. After reaching the preset number of iterations, refresh the actual vehicle location, actual UAV service status, current visibility status of low-orbit satellites, and remaining node resources before configuration distribution. First, reconfirm that the selected configuration's two-level path still meets the current conditions. Then, recalculate resource allocation under fixed paths and task ratios, and normalize according to the preset strict tolerance check ratio, active resources, bandwidth capacity, computing capacity, and UAV energy conditions. Only when all checks pass simultaneously will the final mobile coverage risk be recalculated using the updated vehicle predicted location and the future location extrapolated from the UAV's actual service location, current speed, and current heading, and an executable configuration be generated. If any check fails, a rollback configuration for all local processing will be generated.

[0110] Step 203-2: Deploy the overall configuration or roll back all local configurations. The executable configuration will synchronously deploy the two-level association, three-branch task ratios, link bandwidth, and node CPU resources as a complete configuration unit, along with a final mobile coverage risk assessment after successful verification. Local branches are enabled when the local ratio is greater than zero; vehicle-to-drone uploads are executed only once when the sum of the two remote ratios is greater than zero; drone calculations are enabled when the drone processing ratio is greater than zero; relay and satellite calculations are enabled when the LEO satellite processing ratio is greater than zero and the path is currently available. In the event of a rollback, the local processing ratio is set to one, and the two-level association, drone and LEO satellite processing ratios, remote link bandwidth, and remote CPU resources are all set to zero; incurred flight energy consumption is included in the energy budget.

[0111] Step 203-3: Status and Energy Budget Update. After completing the overall deployment or rollback, update the vehicle history window, UAV service status, energy budget, and LEO satellite status. The energy balance is recursively calculated based on actual consumption and used as input for the feasibility check in the next control cycle. In subsequent cycles, the current link feasible domain is regenerated for newly arriving missions, and the above decisions are repeated.

[0112] The trajectory control strategy described above is trained before deployment. Each training cycle performs prediction, action generation, and feasibility processing, and runs the lower-level decision-making steps 202 to 203 to obtain a shared reward. .in The weighted cost is calculated based on the configuration actually executed in the current control cycle, including normalized task completion delay, normalized system energy consumption, and deadline violation. This cost is applied when resource recalculation and verification pass. Based on the final uninstallation and resource configuration calculations, when verification fails and a full local rollback is executed, Calculated based on all local processing configurations; This represents the normalized mean of the proximity between drone pairs when they are below the warning distance. This serves as a flag indicating that the final resource solution verification failed. and The weights are non-negative.

[0113] During training, interaction sequences are collected using a fixed old policy, and generalized advantage estimation is used to calculate and normalize the advantage. Value bootstrapping is stopped upon true termination, but retained when the time cap is truncated. Each independent policy network is updated according to the pruning probability of the near-end policy optimization compared to the target, while the shared value network is updated according to the value target. After training, the policy checkpoint with the lowest average cost is selected from the retained validation scenarios. During the execution phase, a centralized value network is not required to participate in action generation.

[0114] As a training parameter configuration, the discount factor is set to 0.99, the generalized advantage estimation parameter to 0.95, the pruning threshold to 0.2, the interaction sequence length to 64, the mini-batch size to 128, and each batch of data is updated 10 times; the Adam learning rate for the policy network and the value network is set to 3×10⁻. 4 The entropy regularization coefficient is 0.001, and the above parameters can be adjusted according to training stability.

[0115] In one embodiment of the present invention, 20 vehicles, 4 drones, and 2 low-orbit satellites can be deployed in a 2000m×2000m area, with a control cycle of 1 second. The vehicle speed range is 5–25 m / s, the drone altitude is 100m, the maximum speed is 30 m / s, the maximum three-dimensional service distance is 800m, the minimum drone spacing is 50m, and the initial drone battery power is 100kJ. Each drone has an access bandwidth capacity of 20MHz, each satellite has a bandwidth capacity of 100MHz, and the CPU capacities of the vehicles, drones, and satellites are 1GHz, 5GHz, and 50GHz, respectively.

[0116] In this embodiment, the satellite altitude is set to 550 km, the minimum elevation angle to 11°, the task input data volume can be 1–5 Mbit, the computation workload can be 0.3–8 Gcycles, and the deadline can be 2–6 seconds; the cost weight α = 0.5. =6s, =300J. A lower-level search during the execution phase can be performed using 15 particles and 12 rounds of updates. During the training phase, to reduce training overhead, a 3-particle, 2-round update configuration can be used. The above parameters are an example configuration; the node size and values ​​can be adjusted according to the application scenario.

[0117] For example, if the vehicle's current location is within the drone's coverage area and the drone has currently visible low-Earth orbit (LEO) satellites, all three processing branches—local, drone, and satellite—can be considered for candidate configurations. If the drone has no currently visible LEO satellites, the satellite processing branch is disabled. If the vehicle has no connected drones, only the local branch is retained. When both drones are within the current link's feasible domain, their future positions are extrapolated based on their current speed and heading, and the mobile coverage risk is calculated along with the vehicle's multi-step predicted position. Subsequently, continuous resources are solved for the candidate configurations, and candidates that fail to solve for resources are eliminated. The selected configuration undergoes recalculation and verification before deployment; if verification passes, the risk is recalculated; if verification fails, a full local rollback is performed, and the remote decision and resources are reset to zero.

[0118] Example 2

[0119] To verify the role of mobility prediction information under coverage switching conditions, while keeping the trajectory controller, candidate generation rules and resource solution budget consistent, 12 vehicles, 4 UAVs and 12 control cycles were set up, and 5 scene seeds were used for comparative simulation. The results are shown in Table 1.

[0120] Table 1. Simulation results of motion prediction (mean of seeds for 5 scenarios)

[0121]

[0122] Table 1 shows that, under the same conditions, using motion prediction reduces the average weighted cost by 15.83%, increases the task completion rate within the deadline by 5.28 percentage points, reduces the task interruption rate by 3.61 percentage points, and decreases the average number of retransmissions per scenario by 1.0. The results are the average of the five scenario seeds, and the specific improvement is related to vehicle load and motion trajectory.

[0123] Furthermore, in benchmark validation with 20 vehicles, 4 drones, 100 control cycles, and 10 test scenario seeds, the complete method achieved an average task completion rate of 95.56%, an average system energy consumption of 166.29 J per control cycle, a resource solution failure rate and a resource rollback rate of 0, and a maximum constraint violation of 2.22 × 10⁻¹. 6 The maximum data conservation error is 2.99 × 10⁻¹ 6 The maximum constraint violation is taken as the maximum value among the positive parts of the constraint violation values ​​in the continuous resource allocation problem and the normalized UAV energy violation value; the maximum data conservation error is taken as the maximum value of the relative error between the sum of the data amounts of the three processing branches of each task and the corresponding data amount arriving at the task. The results show that the output configuration satisfies the bandwidth, computation, and data conservation constraints under the given parameters.

[0124] The node size, prediction window length, network capacity, trajectory, and candidate search algorithm can all be adjusted according to the application scenario. This method completes feasible domain establishment, risk calculation, resource screening, and recalculation verification according to the control cycle, and then issues or rolls back the entire system accordingly.

[0125] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks facing mobile coverage risks, characterized in that, This air-space-ground vehicle-to-everything (V2X) task offloading and resource allocation method for addressing mobile coverage risks includes the following steps in each control cycle: Step 1: Generate the current link feasible domain: Collect continuous historical position, speed, and heading of vehicle terminals and current computing task information, and obtain UAV position, speed, heading, remaining energy, low-orbit satellite position, and current visibility status to generate a vehicle position sequence containing multiple predicted times; generate UAV candidate motions based on the predicted position at the last time moment in the vehicle position sequence and the current network status, and determine the current service status of the UAV after checking the service area, UAV spacing, and flight energy; form the current link feasible domain based on the vehicle-UAV path that meets the current coverage conditions between the current vehicle position and the current service position of the UAV, and the UAV-satellite path that meets the current visibility conditions between the current service position of the UAV and the currently visible low-orbit satellite, and freeze the current link feasible domain during the candidate configuration evaluation period of this control cycle; Step 2: Sequentially execute candidate configuration decoding, mobile coverage risk calculation, and resource feasibility solving: Within the current link feasible domain, establish candidate configuration codes that include vehicle-UAV selection, UAV-satellite selection, and the processing ratios of local, UAV, and low-Earth orbit satellite tasks, and mask unexecutable remote processing branches based on the current link feasible domain; for each decoded candidate configuration, using the current service location of the UAV determined in Step 1 as the extrapolation starting point, obtain the future positions of the UAV at multiple predicted times according to the current speed and current heading of the UAV, and calculate the relationship between the predicted vehicle position and the future UAV position at the same predicted time. The service range out-of-bounds degree is used to calculate the mobile coverage risk, and it is weighted according to the sum of the processing ratio of UAVs and the processing ratio of low-orbit satellites. Then, the two-level path relationship and task processing ratio are fixed to solve the link bandwidth and node computing resources that meet the current link capacity, node computing power and UAV energy constraints. Candidate configurations that have not obtained resource feasible solutions are eliminated. Only for candidate configurations that have obtained resource feasible solutions, the resource configuration cost and the mobile coverage risk are combined for ranking. The vehicle predicted location and mobile coverage risk do not add, replace or restore the path in the current link feasible domain, and do not allow resource infeasible candidates to enter the ranking. Step 3: Execution Recalculation and Overall Configuration: Select the candidate configurations corresponding to the sorting results in Step 2. Re-verify the two-level paths using the updated vehicle actual location, UAV actual service status, current visibility status of LEO satellites, and remaining node resources before configuration distribution. Resolve the link bandwidth and node computing resources under the condition of fixed two-level path relationships and task processing ratios. Only when the two-level paths, task processing ratios, activity resources, node capacity, and UAV energy constraints pass the verification simultaneously, recalculate the final mobile coverage risk according to the updated vehicle predicted location and the UAV future location extrapolated from the UAV actual service location, current speed, and current heading. Simultaneously distribute the association relationships, task processing ratios, link bandwidth, and node computing resources as complete configuration units. If any verification fails, set the current task to be processed entirely locally, and set the vehicle-UAV association, UAV-satellite association, UAV and LEO satellite processing ratios, and corresponding remote bandwidth and computing resources to zero. At the same time, retain the UAV flight energy consumption that has occurred, and update the vehicle historical status and UAV energy budget.

2. The method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risks, as described in claim 1, is characterized in that... In step one, the vehicle position sequence is generated by a residual sequence-to-sequence gated cyclic unit model, which includes an encoder and a decoder. The encoder encodes the continuous historical state sequence into a hidden state, and the decoder is initialized with the hidden state. In the first prediction step, the current vehicle position is input, and in subsequent prediction steps, the predicted position of the previous prediction step is input. The residuals output by each prediction step are superimposed on the constant velocity extrapolation position of the corresponding prediction step. The predicted position at the last moment of the vehicle position sequence is used as the input of the candidate motion of the UAV, and the complete predicted position sequence is used for the candidate configuration movement coverage risk calculation in step two.

3. A method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risks, as described in claim 1, is characterized in that... In step one, the candidate motions of the UAV include acceleration and heading increments. Based on the candidate motions, the candidate service states are sequentially checked against the service area boundary, the minimum distance between UAVs, and whether the flight energy consumption in this cycle does not exceed the remaining energy. Only the candidate service states that pass the checks are determined as the current service state of the UAV, and the current link feasible domain is generated at once. In step two, when calculating the mobile coverage risk, the future position sequence of the UAV is extrapolated from the current service state according to the current speed and current heading, but the future position sequence does not change the membership of the current link feasible domain in this control cycle.

4. A method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risks, as described in claim 1, is characterized in that... In step two, for K vehicles and M drones, the candidate configuration code has 3K and M normalized components. The K components select the drones associated with the vehicles, the M components select the low-orbit satellites associated with the drones, and the other 2K components determine the logarithmic weight of the task split for vehicle local processing and drone processing. The logarithmic weight of the task split for low-orbit satellite processing is fixed to zero. Each associated selection component is mapped only to the candidate node index in the current link feasible domain; link bandwidth and node computing resources are not used as candidate configuration coding components.

5. A method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risks, as described in claim 4, is characterized in that... The task components are mapped from [0,1] to the logarithmic weight range of [-8,8]. The task processing ratio is obtained by using softmax normalization with availability mask for the local, UAV and low-Earth orbit satellite processing branches. The local processing branch is always available. The drone and low-Earth orbit satellite processing branches are disabled when there is no vehicle-drone path in the current link feasible domain. The low-Earth orbit satellite processing branch is disabled when there is no selected drone-satellite path in the current link feasible domain. The task processing ratio below the preset threshold is set to zero and then renormalized, and the association is activated only when the corresponding remote task processing ratio is greater than zero.

6. A method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risks, as described in claim 1, is characterized in that... In step two, the mobility coverage risk is first calculated for the decoded candidate configurations. Then, under the condition of fixed two-level path relationships and task processing ratios, the resource allocation problem is solved with link bandwidth and node computing resources as continuous variables. Only when a finite solution is obtained that satisfies the constraints of positive bandwidth of active links, positive computing resources of active processing branches, total node capacity, and UAV remaining energy, is the candidate configuration added to the resource feasible candidate set, and the resource allocation cost is merged with the calculated mobility coverage risk to form a fitness. Candidate configurations not included in the resource feasibility candidate set are not accepted and will not be re-entered for sorting due to low mobile coverage risk values.

7. A method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risks, as described in claim 4, is characterized in that... In step two, starting from the current service position of UAV m determined in step one, and according to the current speed and current heading of the UAV, constant speed and constant heading are extrapolated at control cycle intervals to obtain the future position of UAV m in the h-th prediction step. Calculate the link prediction coverage risk using the three-dimensional distance between the position of vehicle k at the h-th prediction step and the future position of the UAV at the same prediction step, the maximum service distance of the UAV, and the number of prediction steps H: ; in, Let represent the three-dimensional distance between the predicted position of vehicle k at the h-th prediction time and the extrapolated future position of drone m. This indicates the fixed coverage radius preset by the drone m. =max{x,0}; Associating vehicles with drones and remote processing ratio = + Weighted prediction of candidate configuration coverage risk: ; in, This variable indicates the association between vehicle k and drone m; it takes the value 1 when the two are associated, and 0 otherwise. and These represent the proportions of processing by drones and satellites, respectively; all tasks are executed locally. =0; For candidate configurations that satisfy resource and energy constraints, G+ As the fitness, G is the weighted cost obtained from convex optimization. Assign positive infinite fitness to candidate resource-infeasible configurations with a predicted risk weight greater than zero.

8. A method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risks, as described in claim 1, is characterized in that... In step two, the resource allocation cost is obtained by weighting the task completion delay, system energy consumption, and deadline timeout. The deadline timeout is the positive part of the difference between the task completion delay and the task deadline. The task completion delay is the maximum value of the delay of each active processing branch among the local processing branch, the UAV processing branch, and the low-orbit satellite processing branch. The UAV processing branch and the low-orbit satellite processing branch share the same vehicle-UAV upload, and the uploaded data volume is the product of the sum of the processing ratios of the two remote tasks and the task input data volume. The UAV will aggregate and forward the satellite data to be processed associated with the vehicle. The latency of the low-orbit satellite processing branch includes the shared upload latency, the aggregation and forwarding latency of the UAVs passing through, the UAV-satellite propagation latency, and the satellite computing latency. The system energy consumption is the sum of the vehicle local computing energy consumption, vehicle-UAV upload energy consumption, UAV computing energy consumption, UAV forwarding energy consumption to low-orbit satellite, UAV flight energy consumption, and low-orbit satellite computing energy consumption generated during the completion of the vehicle task in the current control cycle. Among them, the vehicle-UAV shared upload energy consumption and aggregation and forwarding energy consumption are each counted only once.

9. A method for task offloading and resource allocation in air-to-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risks, as described in claim 8, is characterized in that... In step two, the continuous resource allocation problem uses the bandwidth of each active link, the computing resources of each active remote processing branch, the task completion delay auxiliary variable, and the deadline timeout auxiliary variable as optimization variables. The task completion delay auxiliary variable is set to be no less than the delay of each active processing branch of the vehicle, and the deadline timeout auxiliary variable is set to be no less than the difference between the task completion delay and the task deadline, and no less than zero. The continuous resource allocation problem satisfies the bandwidth and computing resource capacity constraints of each UAV and low-orbit satellite, the positive resource conditions of the active branches, and the constraint that the sum of the UAV's flight, computing, and forwarding energy consumption does not exceed the remaining energy. The resource allocation for inactive links and processing branches is zero. The resource allocation satisfies the following: ; ; ; ; ; in , For the bandwidth capacity of drones and satellites, This represents the bandwidth allocated to the link from vehicle k to drone m. This represents the bandwidth allocated to the link between the UAV m and the low-Earth orbit satellite s. The variable indicating the association between vehicle k and drone m is... The variable indicating the association between the drone m and the low-Earth orbit satellite s is... This represents the computing resources allocated to drone m for the drone processing branch of vehicle k. This represents the computing resources allocated by low-Earth orbit satellite s to the low-Earth orbit satellite processing branch of vehicle k. , This refers to the CPU capacity of the corresponding node. , and These represent the flight energy consumption, computing energy consumption, and data forwarding energy consumption of UAV m during the current control cycle, respectively. This represents the remaining energy budget of drone m at the start of the current control cycle.

10. A method for task offloading and resource allocation in air-space-ground vehicle-to-everything (V2X) networks oriented towards mobile coverage risks, as described in claim 1, characterized in that... In step two, during the initialization of the candidate configuration search, all locally processed reserved configurations are set. The updated candidate configurations are always re-decoded based on the frozen current link feasible domain, and the mobile coverage risk calculation and continuous resource solution are executed sequentially. In step three, the candidate configurations selected from the sorting results in step two are used as the configurations to be issued. The configurations to be issued are only recalculated based on the updated vehicle predicted position and the future position of the UAV extrapolated according to the current speed and heading of the UAV after all resource recalculation and strict numerical tolerance verification have passed before the configuration is issued. The final mobile coverage risk is then recalculated and issued as a whole. If any verification fails, the local processing ratio of the vehicle is set to 1, the processing ratio of the drone and low-orbit satellite is set to zero, and the two-level association and all remote resources are set to zero. The drone flight energy consumption already included in the energy budget remains unchanged.