Unmanned aerial vehicle task offloading and trajectory planning method and system

Abstract

The application provides a kind of unmanned aerial vehicle task unloading and trajectory planning method and system, it is related to network edge computing technical field, the application includes obtaining task request end state and task information and predicting position and remaining coverage time;Based on predicted position, residual coverage time and task execution risk, generate task risk value;Based on task risk value and predicted position, generate space risk potential field to control the next time position of aerial service node, realize the trajectory planning of unmanned aerial vehicle and other aerial service nodes;Under the updated service condition, first determine the target content source for executable subtask, then determine the target computing node based on the total completion cost including content acquisition cost;End-to-end reliability check is carried out on execution path, and resource is locked after meeting the conditions.The application improves service continuity, task execution reliability and resource utilization efficiency in high dynamic scenario.

Description

Technical Field

[0001] This invention relates to the field of network edge computing technology, specifically to a method and system for unmanned aerial vehicle (UAV) mission unloading and trajectory planning. Background Technology

[0002] With the development of autonomous driving, vehicle-to-everything (V2X) communication, and in-vehicle augmented reality, the computing tasks generated by connected vehicle terminals are characterized by high frequency, heterogeneity, low latency, and strong privacy dependence. To alleviate the computing pressure on terminals, building an air-ground collaborative edge computing system using roadside units, mobile edge computing nodes, and drones has become a mainstream technological direction. Among them, drones, with their high mobility, flexible deployment, and adjustable coverage, have become important aerial service nodes in air-ground collaborative edge computing, providing coverage and computing power support for ground terminals. In existing technologies, task offloading decisions are usually based on the terminal's current location or instantaneous link status to select the optimal service node, and drone trajectory planning is mostly based on the terminal's real-time distribution density or average distance for follow-along movement.

[0003] However, in highly dynamic scenarios such as merging and diverging zones on urban expressways, the high-speed movement of terminals causes them to frequently cross service coverage boundaries. Existing solutions only focus on the connection quality at the current moment, failing to detect the terminal's future tendency to leave the coverage area. This results in the risk of disconnection as soon as the drone establishes a service connection with the terminal, leading to mission interruption or frequent switching. Furthermore, for complex tasks with dependencies, existing technologies often decouple computing node selection from content acquisition, selecting computing nodes first and then searching for the required content. This easily leads to repeated content backhauls across nodes, overlapping subtasks, and wasted link resources. In addition, existing privacy protection mechanisms are mostly independent of the scheduling loop, making it difficult to achieve coordinated optimization of service continuity and resource utilization while ensuring data privacy.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for unloading and planning unmanned aerial vehicle (UAV) missions to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for unmanned aerial vehicle (UAV) mission unloading and trajectory planning includes: Obtain the status information and task information of the task requester, predict the predicted location of the task requester based on the status information, and determine the remaining coverage time of the task requester relative to the current service air node. Based on the predicted location, remaining coverage time, and task information, generate a task risk value corresponding to the task execution risk. A spatial risk potential field is generated based on the mission risk value and the predicted location, and the next moment's location of the air service node is controlled according to the spatial risk potential field. Under the service conditions formed by the updated air service node location, for the executable subtasks in the task information, first determine the target content source, and then determine the target computing node from the candidate computing nodes based on the total completion cost including the content acquisition cost from the target content source to the candidate computing node. Perform end-to-end reliability verification on the execution path formed from the target content source to the target computing node and from the task request end to the target computing node, and lock communication resources and computing resources after the execution path meets the reliability conditions.

[0007] Furthermore, based on the predicted location, remaining coverage time, and task information corresponding to the task execution risk, a task risk value is generated, including: Based on the completion status of subtasks in the task information, determine the set of currently executable subtasks from the task information; For each currently executable subtask in the current set of executable subtasks, estimate the minimum remaining processing time; Based on the minimum remaining processing time and the latest completion time of the currently executable subtasks, generate task deadline risks; Coverage handover risk is generated based on the remaining coverage time, where the shorter the remaining coverage time, the higher the coverage handover risk. Generate link interruption risk based on the link quality between the task requester and the current service air node; Queue waiting risk is generated based on the queue waiting time of candidate computing nodes; Generate cache mismatch risk based on the cache hit status of the content required by the currently executable subtask on the candidate content source; The risk values ​​of task deadline, coverage switching, link interruption, queue waiting, and cache mismatch are weighted and fused to obtain the task risk value.

[0008] Furthermore, based on the minimum remaining processing time and the latest completion deadline of the currently executable subtasks, task deadline risks are generated, including: Local processing nodes, service nodes, air service nodes, and ground edge nodes are selected as candidate execution nodes; For each candidate execution node, the estimated processing time of the current executable subtask on the candidate execution node is determined based on the available transmission rate from the task requester to the candidate execution node, the amount of data of the current executable subtask, the allocable computing frequency of the candidate execution node, the amount of computing required by the current executable subtask, and the queue waiting time of the candidate execution node. Select the minimum value from the estimated processing time corresponding to the candidate execution nodes as the minimum remaining processing time; Calculate the ratio between the minimum remaining processing time and the remaining time limit of the currently executable subtask, where the remaining time limit is the difference between the latest completion time limit and the current time. The largest ratio in the current set of executable subtasks is determined as the task deadline risk; Based on the remaining coverage time, coverage handover risks are generated, including: Compare the remaining coverage time with the preset prediction time window; Increase the risk of coverage handover when the remaining coverage time is less than the preset prediction time window; Reduce the risk of coverage handover when the remaining coverage time is greater than or equal to the preset prediction time window; Based on the link quality between the task requester and the current service air node, link interruption risks are generated, including: Obtain the signal-to-noise ratio or average signal-to-noise ratio between the task requester and the current service's air service node; Increase the risk of link interruption when the signal-to-noise ratio or average signal-to-noise ratio is lower than or equal to a preset interruption threshold; Reduce the risk of link interruption when the signal-to-noise ratio or average signal-to-noise ratio is higher than the preset interruption threshold; Queue waiting risk is generated based on the queue waiting time of candidate computing nodes, including: Get the current queue waiting time of the candidate computing node; Increase the risk of queue waiting when the queue waiting time increases relative to the previous detection time; Reduce queue waiting risk when the queue waiting time decreases relative to the previous detection time; Based on the cache hit status of the content required by the currently executable subtask on the candidate content source, a cache mismatch risk is generated, including: Determine whether the content required by the currently executable subtask is cached in at least one candidate content source; Increase the risk of cache mismatch if the required content is not cached in any of the candidate content sources; Reduce the risk of cache mismatch when the required content is cached in at least one candidate content source.

[0009] Furthermore, a spatial risk potential field is generated based on the mission risk value and predicted location, and the next-time location of the airborne service node is controlled according to the spatial risk potential field, including: The predicted location of the task requester is used as the risk center; Multiple spatial location points are set within the service area plane; Based on the task risk value and the distance between the spatial location point and the predicted location, the risk potential value of each spatial location point is determined. The higher the task risk value, the greater the risk potential value; the closer the spatial location point is to the predicted location, the greater the risk potential value. The direction of movement of the air service node is determined based on the increasing risk potential at its current location; Update the planar position of the air service node according to the direction of movement to obtain the next position of the air service node.

[0010] Furthermore, the planar position of the air service node is updated according to the direction of movement to obtain the next position of the air service node, including: Obtain the gradient direction of the risk potential field at the current location of the air service node; Move along the gradient direction of the risk potential field by a preset step size to obtain the initial position for the next time step. Obtain the location of other air service nodes; Based on the distance between the location of other air service nodes and the current location of the air service node, a repulsion displacement is generated. The smaller the distance between the current location and other air service nodes, the larger the repulsion displacement. Based on the initial next-time position and the repulsion displacement, candidate next-time positions are generated; Determine whether the displacement length between the candidate next position and the current position of the air service node exceeds the preset maximum flight distance; If the displacement length exceeds the preset maximum flight distance, the displacement vector of the candidate next moment position relative to the current position is scaled proportionally according to the ratio of the preset maximum flight distance to the displacement length to obtain the next moment position. If the displacement length does not exceed the preset maximum flight distance, the candidate position for the next moment is determined as the position for the next moment.

[0011] Furthermore, for the executable subtasks in the task information, the target content source is first determined, and then, based on the total completion cost including the content acquisition cost from the target content source to the candidate computing nodes, the target computing node is determined from the candidate computing nodes, including: For each executable subtask in the task information, determine the required content for the executable subtask; The candidate content source set is determined from the task request end, service node, air service node, ground edge node and cloud, and the candidate computing node set is determined from the service node, air service node, ground edge node and cloud; For each candidate content source in the candidate content source set and each candidate computing node in the candidate computing node set, determine the content acquisition cost for the candidate content source to provide the required content to the candidate computing node; The candidate content source with the lowest content acquisition cost is identified as the target content source; After the target content source is determined, for each candidate computing node in the candidate computing node set, the total completion cost of the candidate computing node is determined. The total completion cost includes the transmission cost of the task requester uploading subtask data to the candidate computing node, the queuing cost of the candidate computing node, the computing cost of the candidate computing node, and the content acquisition cost of the target content source providing the required content to the candidate computing node. The candidate computing node with the lowest total completion cost is selected as the target computing node.

[0012] Furthermore, the total cost of completion also includes energy costs, dependency penalties, and privacy penalties; Determine the total completion cost of candidate computing nodes, including: The latency cost is determined based on the data upload time from the task requester to the candidate computing node, the queue waiting time of the candidate computing node, the computing time of the candidate computing node for the executable subtask, and the content acquisition time of the target content source to transmit the required content to the candidate computing node. The energy cost is determined based on the transmission energy consumption generated by the task requester uploading executable subtask data and the computing energy consumption generated by the candidate computing nodes executing executable subtasks. Get the completion time of the predecessor subtask of the executable subtask; When the completion time of the predecessor subtask is later than the planned start time of the executable subtask on the candidate computing node, the dependency penalty is determined based on the difference between the completion time and the planned start time of the predecessor subtask. The larger the difference, the larger the dependency penalty. When the candidate computing node is not local to the task requester, the privacy penalty is determined based on the privacy level of the executable subtask; the higher the privacy level, the greater the privacy penalty. The total completion cost is determined by a weighted sum of latency cost, energy cost, dependency penalty, and privacy penalty.

[0013] Furthermore, end-to-end reliability verification is performed on the execution paths formed from the target content source to the target computing node and from the task requester to the target computing node. Communication and computing resources are locked once the execution path meets the reliability requirements, including: The link between the task requester and the target computing node is defined as the uplink segment. The link between the target content source and the target computing node is defined as the content transmission link segment; Obtain the probability of uplink segment interruption and the probability of content transmission segment interruption; The end-to-end interruption probability is determined based on the interruption probability of the uplink segment and the interruption probability of the content transmission segment. The higher the interruption probability of the uplink segment, the higher the interruption probability of the content transmission segment, and the higher the end-to-end interruption probability. Determine whether the end-to-end interruption probability does not exceed a preset reliability threshold; If the end-to-end interruption probability does not exceed the preset reliability threshold, the target computing node is retained, and the computing frequency resources and link bandwidth resources of the target computing node are locked. If the end-to-end interruption probability exceeds a preset reliability threshold, the target computing node is removed from the candidate computing node set, and the process of determining the target content source and target computing node is returned until the end-to-end interruption probability corresponding to the re-determined execution path does not exceed the preset reliability threshold.

[0014] Furthermore, after locking the computing frequency resources and link bandwidth resources of the target computing node, it also includes: Record the difference between scheduling features and actual costs. Scheduling features include at least one of the following: data volume of executable subtasks, computational volume, privacy level, speed of task requesting end, heading, task risk value, remaining coverage time, link interruption probability, queue waiting time, and content cache hit status. The actual cost difference is the difference between the actual execution cost and the predicted cost. Local correction parameters are generated based on the difference between scheduling characteristics and actual costs. After the local correction parameters are generated, a preset random backoff time is allowed before uploading the local correction parameters to the cloud. Receive global correction parameters from the cloud side. The global correction parameters are obtained by the cloud side by weighting and aggregating the local correction parameters based on the number of samples from the task requesting end that uploaded the local correction parameters. Update the minimum remaining processing time estimate and total completion cost estimate in the next control cycle using global correction parameters; After updating the minimum remaining processing time estimate and total completion cost estimate for the next control cycle using the global correction parameters, the following steps are also included: In the next control cycle, based on the updated air service node location, link status, queue status, cache status and global correction parameters, the process of predicting location, task risk value, spatial risk potential field, target content source, target computing node, end-to-end reliability verification and locking of communication resources and computing resources will be re-executed. Obtain the comprehensive optimization target, which is determined based on subtask completion latency, energy consumption, end-to-end interruption probability, and privacy penalty. Determine whether all executable subtasks have been completed; If all executable subtasks are not completed and the overall optimization target decreases by a factor greater than or equal to a preset threshold relative to the previous control cycle, the next control cycle will continue to be executed. Task scheduling ends when all executable subtasks are completed, or when the overall optimization target decreases less than a preset threshold relative to the previous control cycle.

[0015] This invention further provides a UAV mission unloading and trajectory planning system, the system being used to implement the above-mentioned UAV mission unloading and trajectory planning method, comprising: The data acquisition module is used to acquire the status information and task information of the task requesting end, predict the predicted location of the task requesting end based on the status information, and determine the remaining coverage time of the task requesting end relative to the current service air node. The risk assessment module is used to generate task risk values ​​based on the predicted location, remaining coverage time, and task information corresponding to the task execution risk. The location prediction module is used to generate a spatial risk potential field based on the mission risk value and the predicted location, and to control the next moment's location of the air service node according to the spatial risk potential field. The target determination module is used to determine the target content source for executable subtasks in the task information under the service conditions formed by the updated air service node locations, and then determine the target computing node from the candidate computing nodes based on the total completion cost including the content acquisition cost from the target content source to the candidate computing node. The resource verification module is used to perform end-to-end reliability verification on the execution path formed from the target content source to the target computing node and from the task request end to the target computing node, and lock communication resources and computing resources after the execution path meets the reliability conditions.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention predicts the location of a task requester based on its state information and determines the remaining coverage time of the task requester relative to the current service node. Then, based on the predicted location, remaining coverage time, and task information, it generates a task risk value corresponding to the task execution risk. A spatial risk potential field is generated based on the task risk value and the predicted location. The spatial risk potential field is used to control the next position of the service node, causing the service node to move towards areas with higher future task execution risks. This allows the service node to reach areas prone to task timeouts and coverage loss in advance, shortening the service reconstruction time when the task requester crosses coverage boundaries and improving service continuity in highly dynamic scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method flow of the present invention.

[0018] Figure 2 This is a schematic diagram of the overall system structure of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Example: Please see Figure 1 This embodiment provides a method for unmanned aerial vehicle (UAV) mission offloading and trajectory planning, which is applied to the scheduler or airborne service node itself in an air-ground collaborative edge computing system. The air-ground collaborative edge computing system adopts a four-layer collaborative structure of terminal, air, ground, and cloud, including a mission request terminal, airborne service node, ground edge node, service node, and cloud server.

[0022] Specifically, the task requester is the in-vehicle terminal that generates the computational task. In highly dynamic vehicle-to-everything (V2X) scenarios such as merging and diverging zones on urban expressways, the task requester runs applications such as autonomous driving assistance, vehicle-to-infrastructure (V2I) cooperative warnings, and in-vehicle visual analysis. The application tasks it generates are decomposed into directed acyclic graph (DAG) tasks with dependencies. The task requester undertakes three responsibilities: First, it reports its own status and task information, providing input for location prediction and risk assessment; second, as one of the candidate execution nodes, it can locally execute sub-tasks with low data sensitivity; third, based on local samples, it trains a lightweight correction model locally and only uploads the locally corrected parameters, without uploading the original trajectory, original task content, and perception data, thus protecting data privacy while supporting continuous correction of cost estimation.

[0023] Aerial service nodes serve as mobile edge computing platforms such as drones, and their coverage areas can proactively move according to control commands. When it is predicted that a task requesting an end in a certain area will face task timeout or coverage loss, the aerial service node moves to that area in advance, driven by the spatial risk potential field, providing flexible coverage and computing power support. Simultaneously, the aerial service node also participates in the selection of target computing nodes and target content sources for subtasks, acting as one of the candidate computing nodes and candidate content sources.

[0024] Ground edge nodes are fixed roadside units or mobile edge computing nodes with a fixed coverage radius, serving as the normal service providers for task requesters. Task requesters may cross the coverage boundaries of ground edge nodes during high-speed movement, which is the primary source of service interruption risk; the quantification of remaining coverage time and coverage switching risk is based on this. Service nodes are nodes with computing and caching capabilities, such as service vehicles, which, together with ground edge nodes, constitute the ground support layer. Both participate in scheduling selection as candidate execution nodes and candidate content sources.

[0025] The cloud server performs two functions: First, global parameter aggregation and model updating, which means receiving local correction parameters asynchronously uploaded by each task requesting end, weighting and aggregating them according to the number of samples from each task requesting end to obtain global correction parameters and distributing them to each node, so that the system as a whole can continuously correct cost estimates without the original data leaving the local area; Second, as a remote content source, when the content required by the subtask is not cached in any edge candidate content source, the required content is returned by the cloud server.

[0026] In terms of connectivity, the task requester maintains a primary service connection with the airborne service node via an uplink link, with link quality measured by indicators such as signal-to-noise ratio and interruption probability. Simultaneously, the task requester is located within the coverage area of ​​the ground edge node. During subtask execution, an execution path is formed, consisting of two links: one from the task requester to the target computing node, and the other from the target content source to the target computing node. Resource locking is only performed after the end-to-end reliability of the execution path has been verified. The interaction between the cloud server and each node is low-frequency and parameter-level. The method provided in this embodiment can be executed by the scheduler in the system or by the airborne service node itself. This embodiment achieves forward-looking deployment and content-computing power collaborative scheduling driven by future risk awareness through the following five steps.

[0027] The specific steps of this embodiment include: Step 1: Obtain the status information and task information of the task requester, predict the predicted location of the task requester based on the status information, and determine the remaining coverage time of the task requester relative to the current service air node.

[0028] Specifically, the scheduler first obtains the status information of the task requester at time t. (t), this state information includes two-dimensional position coordinates (t). (t), (t) and velocity (t), heading angle (t) and acceleration (t). The status information also includes the locally available computing frequency. and local energy consumption coefficient ,in, This indicates the computational frequency that the task requesting end can currently allocate to local task execution. This represents the proportion of energy consumed per unit of computation by the task requesting end. It is determined by the chip's hardware process and circuit architecture, and can be obtained by consulting the chip datasheet or through actual measurement and calibration of typical computational loads. Locally available computing frequency. and local energy consumption coefficient This provides a basis for subsequent estimation of local execution costs. Simultaneously, task information is acquired. Specifically, the application tasks generated by the task requester are decomposed into a directed acyclic graph (DAG). For any application task of task requester i, it is represented as... ,in, A set of subtasks, containing Sub-tasks This is the set of dependent edges. If the subtask... sub-tasks If there are dependency edges, it indicates a subtask. Only in subtasks Execution can only begin after all tasks are completed. Each subtask... Defined as a quadruple ,in, Indicates the amount of input data for the subtask. This represents the number of CPU cycles required to complete the subtask. This indicates the content identifier that the subtask depends on for execution. This represents the latest completion time of the subtask. The decomposition of the directed acyclic graph (DAG) is achieved through static analysis of the application's execution logic: extracting the call relationships and data transfer relationships between various processing stages in the application, identifying each processing stage as a subtask, defining the sequential dependencies of data transfer as dependency edges, and calculating the input data volume, required CPU cycles, dependent content identifiers, and the latest allowed completion time for each subtask. Through this decomposition, the original overall application task is transformed into a set of independently schedulable subtasks with clearly defined dependencies. This allows the scheduler to accurately determine which subtasks currently have the conditions for execution, avoiding invalid resource estimations for subtasks whose predecessors have not yet completed, and providing structured objects for subsequent risk quantification and content-computation joint scheduling. Addressing the rapid changes in the motion state of the task request end in highly dynamic scenarios... This embodiment uses a road-constrained short-time motion prediction model to calculate the predicted position at the future time Δt. The predicted x-coordinate and y-coordinate are calculated using the following formulas: in, and These represent the predicted horizontal and vertical coordinates of the task requester at a future time Δt; and The coordinates of the current position; Current speed; This is the current heading angle; This is the current acceleration; Preset prediction time window. Introduce acceleration. This section aims to correct the error of the uniform linear motion assumption in high-dynamic variable speed scenarios, significantly improving short-term prediction accuracy. For example, in the merging and diverging zone of an urban expressway, the task requester frequently accelerates and decelerates. If the acceleration term is ignored, the predicted position deviation may exceed 50 meters, leading to subsequent misjudgments of risk. It should be noted that the short-term motion prediction model in this embodiment is a kinematic model based on the uniform acceleration motion assumption, rather than a neural network prediction model. The reason for using a kinematic model is that the prediction time window is short, and the motion state of the task requester within this time scale can be approximated by the uniform acceleration assumption; moreover, the kinematic model only involves arithmetic operations and trigonometric function operations, with a small amount of computation per prediction and stable output latency, which can meet the requirements of real-time calling on a control cycle basis; while neural network prediction models require a large number of historical trajectory samples for training, and the inference latency fluctuates, making it difficult to guarantee stable output on a cycle basis. The road constraint is reflected in the fact that the prediction result can be further projected onto the road centerline to eliminate prediction components that deviate from the road direction. This projection operation can be omitted when the task requester is traveling along a straight road. Preset prediction time window The prediction time window is set based on the ratio between the maximum speed of the requesting client and the coverage radius of the current air service node, for example, between 0.5 and 2.0 seconds. The rationale for this setting is as follows: if the prediction time window is too short, the predicted location can only reflect the recent movement trend of the requesting client, and the air service node will not have enough time to complete forward deployment; if the prediction time window is too long, the error of the uniform acceleration assumption accumulates over time, and the reliability of the predicted location decreases. Using the ratio of coverage radius to maximum speed as the upper bound of the prediction time window ensures that the critical periods during which the requesting client may cross the coverage boundary are predicted, providing effective look-ahead time for subsequent risk assessment and trajectory control.

[0029] After obtaining the predicted location, the predicted distance between the requesting client and the current service's airborne node is first calculated based on the predicted location. Let the coordinates of the current service's airborne node be... Then predict the distance. Calculated using the following formula: in, and These represent the x and y coordinates of the current service's airborne node, respectively. The meanings of the remaining parameters are consistent with the predicted location calculation formula described above. By comparing the predicted distance with the coverage radius of the current service's airborne node, it can be determined whether the requesting client will still be within the node's coverage area in the future: if the predicted distance is not greater than the coverage radius, it means the requesting client will still be within the coverage area of ​​the current service's airborne node in the future; if the predicted distance is greater than the coverage radius, it means the requesting client is about to leave the coverage of the current service's airborne node, requiring advance service migration or forward deployment of airborne service nodes. The judgment based on the predicted distance shifts the service continuity issue from a passive switch after it occurs to a pre-emptive prediction. Based on the predicted distance, the remaining coverage time of the requesting client relative to the current service's airborne node is further determined.

[0030] Let the coverage radius of the current service air node be... The distance between the task requester and this node at the current moment is Then the remaining coverage time Calculated using the following formula: Among them, molecules Indicates the remaining spatial distance between the task requesting end and the area outside the coverage zone; denominator Indicating the urgency of leaving the coverage area, To prevent extremely small constants with a denominator of zero (e.g., taking...) The physical meaning of this formula is: the smaller the remaining spatial distance, the faster the movement speed, and the shorter the remaining coverage time, the closer the risk of service interruption. This indicator directly quantifies the vulnerability of service continuity in the time dimension, providing a crucial time benchmark for subsequent risk calculations. In the above formula for calculating the remaining coverage time, the distance between the task requesting end and the current service node in the air is... The method uses the same Euclidean distance form as the predicted distance, but instead of the predicted location coordinates, it uses the current location coordinates. The remaining coverage time is approximated by a first-order form of the ratio of the remaining spatial distance to the current velocity. This is reasonable because the remaining coverage time is used for risk prioritization and look-ahead triggering, rather than precise timing. The first-order approximation is simple to calculate and has a fast response, and can be updated with low overhead in each control cycle. At the same time, the numerator uses the remaining spatial distance instead of the predicted distance, which can reflect the true coverage margin at the current moment. Together with the predicted distance judgment, it forms a two-level coverage assessment that combines the current margin with the future trend, improving the robustness of coverage loss prediction.

[0031] Step 2: Generate task risk values ​​based on the predicted location, remaining coverage time, and task information corresponding to the task execution risk.

[0032] This step aims to map multi-dimensional physical states to a unified scalar risk value, characterizing "the extent to which untimely service will lead to system performance degradation." The specific generation process includes the quantification and fusion of the following five sub-dimensions.

[0033] First, based on the completion status of subtasks in the task information, determine the current set of executable subtasks, which are the executable subtasks for which all predecessor subtasks have been completed. Specifically, the current set of executable subtasks... Determined by the following formula: in, Subtasks The set of predecessor subtasks, i.e., based on the set of dependency edges. Definitely required in subtasks Begin executing all previously completed subtasks; This represents the set of subtasks completed by task requester i up to time t. The determination process is as follows: For each subtask, retrieve its predecessor subtask set and determine if the predecessor subtask set is a subset of the completed subtask set. If yes, include the subtask in the current executable subtask set; otherwise, do not include it until its predecessor subtask is completed, at which point it will be re-evaluated in a subsequent control cycle. This determination ensures that only subtasks truly ready for execution enter the subsequent risk assessment and scheduling phases, avoiding ineffective cost estimations and resource reservations for subtasks whose predecessors are not yet completed.

[0034] For each currently executable subtask in the set, estimate the minimum remaining processing time. This time is obtained by traversing candidate execution nodes such as local processing nodes, service nodes, air service nodes, and ground edge nodes, and calculating the minimum value of the sum of transmission latency, computation latency, and queue waiting time, as shown in the following formula: in, For the amount of data in the subtask, For available transmission rate, The required number of CPU cycles. Let j be the allocatable computational frequency. This represents the queue waiting time. The reason for taking the minimum value for each candidate execution node in the above formula is that the minimum remaining processing time describes the fastest completion time that the subtask can achieve on the most favorable execution node under the current resource conditions, representing the optimal service capacity that the system can provide for the subtask. Using the optimal service capacity as the assessment benchmark for deadline risk can objectively reflect the true urgency of task timeout: if even selecting the optimal node cannot meet the time limit requirement, it means that the task is in a high-risk state and must be responded to first; otherwise, it means that the task still has sufficient scheduling margin. The above formula adds the transmission delay, computation delay, and queue waiting time because the processing process of the subtask on any node consists of three serial steps: data upload, queuing, and computation execution. The time consumed by these three steps is directly added together to form the complete processing time. Based on this, the task deadline risk is generated. As shown in the following formula: in, This represents the latest completion deadline for the subtask. The "maximum ratio" is used here instead of the average, based on the "weakest link" principle: the overall completion time of a DAG task depends on the most urgent critical subtask. If even one subtask faces timeout risk, the overall deadline risk of the task should be increased, thus triggering the system's priority response mechanism.

[0035] Secondly, coverage handover risks are generated based on the remaining coverage time. As shown in the following formula: in, As a normalization window, it is used to eliminate the impact of different prediction durations on risk values. An exponential decay form is used instead of a nonlinear function because in real communication systems, when the remaining coverage time approaches zero, the probability of link handover failure increases nonlinearly. The exponential function can keenly capture this "critical collapse" characteristic, enabling the system to rapidly increase the risk weight when the requesting end is about to leave the coverage area, driving the over-the-air service node to intervene in advance.

[0036] Next, link interruption risk is generated based on the link quality between the task requester and the current service's airborne service node. Specifically, the instantaneous signal-to-noise ratio of the link between the task requester and the current service's airborne service node is first obtained. It is calculated using the following formula: in, This indicates the transmit power of the requesting end. This represents the channel gain from the requesting end to the current serving air node, including the effects of path loss and small-scale fading. Indicates noise power; This represents the co-channel interference power, which is the sum of the interference power caused to the current link by other transmission links using the same frequency band at the same time. The above formula expresses link quality as the ratio of useful signal power to the total power of interference plus noise. This ratio directly determines the upper limit of the achievable transmission rate of the link and is a fundamental quantity for judging whether the link can maintain reliable transmission. Subsequently, the instantaneous signal-to-noise ratio is averaged within a sliding time window to obtain the average signal-to-noise ratio. The purpose of averaging is to suppress rapid fluctuations in instantaneous channel measurements, avoid misjudging link outages due to instantaneous deep fading, and ensure that the assessment of link outage risk reflects the statistical characteristics of the channel rather than instantaneous jitter. Based on this, link outage risk is generated according to the average signal-to-noise ratio.

[0037] Specifically, this is achieved by obtaining the signal-to-noise ratio or average signal-to-noise ratio. Combined with preset interrupt threshold Calculate the probability of interruption As a risk of link interruption, it is shown in the following formula: This formula reveals the negative correlation mechanism between the interruption probability and the signal-to-noise ratio (SNR): the lower the SNR, the greater the probability of signal demodulation failure and the higher the risk of interruption. This ensures that the system can detect potential communication interruption risks when channel conditions deteriorate. The above formula uses the cumulative distribution form of an exponential function, based on the fact that in a Rayleigh fading channel, the received signal envelope follows a Rayleigh distribution, and the corresponding instantaneous SNR follows an exponential distribution. The probability that the SNR is below a certain threshold is equal to the cumulative distribution function value of the exponential distribution at that threshold. Therefore, the above formula is an analytical expression of the interruption probability in a Rayleigh fading channel, which can be quickly calculated without Monte Carlo simulation, meeting the requirements of real-time evaluation on a cycle-by-cycle basis. (Preset interruption threshold) The signal-to-noise ratio (SNR) threshold is set based on the minimum required signal-to-noise ratio (SNR) of the modulation and coding scheme used in the link. For example, when the link uses quadrature phase-shift keying (QPSK) modulation and the coding rate is halved, the minimum SNR required to maintain reliable demodulation is approximately 6 dB, which can be set as the interruption threshold. The rationale for this setting is that when the SNR falls below this threshold, the bit error rate rises sharply, and the link cannot maintain effective transmission. Determining an interruption based on this threshold is consistent with the actual link behavior, providing a physical basis for quantifying the risk of link interruption.

[0038] Next, queue waiting risk is generated based on the queue waiting time of candidate computing nodes. As shown in the following formula: in, This is a normalization constant, set to the order of magnitude of the typical business processing time of the system, for example, 1 second. Its function is to map queue waiting times of different magnitudes to the dimensionless (0,1) interval, facilitating weighted fusion with other risk components; when the queue waiting time is much greater than... When the risk value approaches 1, it exhibits saturation characteristics, thus preventing the queue duration of extremely congested nodes from excessively dominating the risk value after fusion. If the value is too small, the risk value will saturate prematurely, making it impossible to distinguish between moderate and severe congestion; if the value is too large, the risk value will be insensitive to queue changes. Therefore, the typical business processing time is used as the calibration benchmark. The reason for taking the minimum value for each candidate computing node in the above formula is that subtasks can choose the candidate computing node with the shortest queue to execute. Therefore, the queue waiting risk should reflect the congestion level of the optimal available computing resources in the system. If even the node with the shortest queue has a long waiting time, it indicates that the overall computing resources of the system are strained, and tasks face a general queuing risk.

[0039] Then, based on the cache hit status of the content required by the currently executable subtask on the candidate content source, a cache mismatch risk is generated. As shown in the following formula: ) in, This is the cache hit indicator (1 for a hit, 0 for a miss). If the required content is not found in any candidate content source, the risk value is 1; otherwise, it is 0. Cache mismatch will cause content to need to be sent back from a remote cloud server, significantly amplifying the backhaul latency and consuming backbone network bandwidth. Therefore, it is considered as an independent risk component.

[0040] Finally, the five risks mentioned above are weighted and combined to obtain the comprehensive task risk value. As shown in the following formula: in, to The weights are non-negative and satisfy the following conditions: Regarding the weighting method, this embodiment adopts a combination of "historical data statistical regression + expert experience matrix". Specifically, firstly, historical operational data is collected, and the correlation coefficients between each risk component and the actual task failure rate are analyzed; then, combined with business priorities (e.g., security-related businesses place more emphasis on cutoff risk and link risk), experts fine-tune the regression results. In a specific example targeting a highly dynamic urban vehicle-to-everything (V2X) scenario, the following is set... This set of weights highlights the dominant role of task deadline risk (0.4), as task timeout is the primary factor leading to a collapse in user experience in this scenario. Meanwhile, coverage switching risk and link interruption risk are given higher weights (0.2 each) to address connection instability caused by high-speed movement. Queue and cache risks have lower weights (0.1 each), as they can usually be mitigated through scheduling and have a relatively mild impact. This differentiated configuration ensures that the risk values ​​truly reflect the core pain points in the current scenario. The rationale for using a linear weighted approach for fusion is as follows: First, after the above formula design, each risk component has been normalized to a dimensionless quantity within the (0,1) interval. Linear weighting maintains the monotonic relationship between the fusion result and each component, meaning that the task risk value will not decrease when any component increases, thus ensuring the stability of the gradient direction of the subsequent spatial risk potential field and preventing the nonlinear fluctuations of the fusion function from causing the movement direction of the airborne service node to oscillate back and forth. Second, the linear weights have clear interpretability; the relative importance of each risk dimension can be directly read from the weight values, facilitating calibration and adjustment in conjunction with business scenarios.

[0041] Step 3: Generate a spatial risk potential field based on the task risk value and predicted location, and control the next moment position of the air service node according to the spatial risk potential field.

[0042] This step transforms the abstract task risk into a physical field in space, driving the aerial service nodes to proactively move towards high-risk areas. Specifically, using the predicted location of the task requester as the risk center, multiple spatial location points z are set within the service area plane. Based on the task risk value... The risk potential value for each spatial location is determined by the distance between the spatial location point and the predicted location. As shown in the following formula: in, Represents Euclidean distance. This is the set of all task requesters that have generated task requests within the current system. Represents the predicted location vector. To prevent singular stable terms. The value of this potential field is typically 1% of the coverage radius of the airborne service node (e.g., 5 meters for a coverage radius of 500 meters). Its purpose is to prevent the denominator from becoming zero and resulting in an infinite potential value when the spatial location coincides with the predicted location, thus ensuring the stability of the numerical calculation. The physical meaning of this potential field is: the closer to a high-risk mission requesting end, and the higher the risk value of that requesting end, the greater the potential value at that point. High-risk targets form a local maximum potential value region around them. The reason for using the inverse square law form of the potential field expression is threefold: First, the potential value increases sharply as the distance decreases, creating a strong attraction to nearby high-risk targets, allowing a limited number of airborne service nodes to prioritize responses to the most urgent risks. Second, this form is continuously differentiable throughout space, allowing the gradient direction to be calculated at any location, ensuring the continuity and smoothness of the airborne service node's trajectory. Third, the potential values ​​of multiple mission requesting ends are summed. When multiple high-risk targets exist simultaneously, the extreme value location of the potential field automatically reflects the weighted sum of the risk values ​​and locations of each target, achieving unified guidance for multiple targets without the need for pairwise coordination.

[0043] Subsequently, based on the increasing risk potential at the current location of the airborne service node, its direction of movement is determined. Specifically, this is achieved by calculating the potential field gradient. To achieve this, the potential gradient is calculated analytically: taking the partial derivative of the spatial coordinates in the above risk potential expression, we obtain the following formula: Where, vector This represents the displacement from the predicted position of task requester i to the spatial position. Since the coefficients are negative, the direction of each term in the gradient points from the spatial position to the predicted position of the corresponding task requester. The magnitude of each term is directly proportional to the task risk value and inversely proportional to a power of the distance. Therefore, the gradient direction is essentially a weighted sum of the directions pointed to by each predicted position. The weights are determined by both the task risk value and the distance; targets that are closer and have higher risk exert a stronger pull on the movement direction. Using analytical gradients instead of numerical differencing avoids truncation errors caused by differentiating step sizes, making the determination of the movement direction more accurate. Furthermore, the computational cost is only linearly related to the number of task requesters, allowing calculations to be completed within a single control cycle, meeting real-time requirements.

[0044] Move along the gradient direction of the risk potential field by a preset step size To obtain the initial position for the next time step. As shown in the following formula: in, The step length should be set to 30% to 60% of the maximum flight distance of the airborne service node within each control cycle, and should be matched with the maximum maneuverability of the airborne service node. If the step length is too large, the airborne service node may overshoot the risk center and oscillate back and forth on both sides of the risk center; if the step length is too small, the airborne service node will not be able to reach high-risk areas in time, weakening the effectiveness of forward deployment. For example, when the maximum speed of the airborne service node is 20 meters per second and the control cycle is 1 second, the step length can be set to 6 to 12 meters. For the gradient normalization stable term, take the smallest positive number (e.g. This is used to prevent division by zero errors when the gradient magnitude is zero, which corresponds to the situation where the airborne service node is exactly at the potential field equilibrium position. This update form is essentially a unit vector movement along the gradient direction by a step length. The significance of normalizing the gradient is that the potential field strength varies greatly at different locations, and the gradient magnitude increases significantly near the risk center. If the gradient vector is directly used as the displacement, the movement speed of the airborne service node will change drastically with its position. After normalization to a unit vector, the displacement length in each control cycle is always equal to the step length, resulting in stable and controllable flight speed, which is easier to match with the dynamic constraints of the airborne service node. To prevent multiple airborne service nodes from clustering in the same high-risk area, leading to increased co-channel interference and service redundancy, a repulsive displacement mechanism is introduced. The repulsive displacement is generated based on the distance between the positions of other airborne service nodes and the current position, as shown in the following equation: in, The repulsion coefficient is set based on simulation calibration results, for example, between 0.1 and 1.0; U represents the set of all airborne service nodes. The calibration principle is to balance two needs: if the repulsion effect is too weak, multiple airborne service nodes will cluster in the same high-risk area, leading to increased co-channel interference and coverage redundancy; if the repulsion effect is too strong, airborne service nodes will be pushed away from high-risk areas, deviating from the risk-driven deployment objective. In the above formula, the repulsion displacement is inversely proportional to the cube of the distance between nodes, that is, when the distance between nodes decreases, the repulsion displacement increases cubically, forcing airborne service nodes to disperse, so that multiple airborne service nodes can cover different high-risk areas while maintaining spatial isolation. On the other hand, if no repulsion term is set, multiple nodes may overlap and hover, causing a sharp drop in the downlink signal-to-interference-plus-noise ratio (SINR), which would instead offset the gain brought by being closer to the user.

[0045] Finally, it is determined whether the displacement length between the candidate next position and the current position exceeds the preset maximum flight distance. If the displacement length exceeds the preset maximum flight distance, the displacement vector of the candidate next moment position relative to the current position is scaled proportionally according to the ratio of the preset maximum flight distance to the displacement length to obtain the next moment position, as shown in the following formula: The use of proportional scaling instead of direct truncation of coordinate components is to maintain consistency in the direction of movement. Direct truncation may cause the actual flight direction to deviate from the potential field gradient direction, preventing the node from accurately approaching the risk center; proportional scaling satisfies both dynamic constraints and preserves the optimal movement intention. It should be noted that during the trajectory update process, the airborne service node maintains a preset flight altitude, and trajectory updates are performed only on its planar position. The preset flight altitude is determined based on building heights and airspace restrictions within the service area, for example, set to 50 to 100 meters. The preset maximum flight distance is the product of the UAV's maximum flight speed and the control cycle. The next position determined sequentially in each control cycle constitutes the flight trajectory of the airborne service node. Through this process, the UAV continuously moves towards areas with higher future mission execution risks, achieving mission risk-oriented trajectory planning.

[0046] Step 4: Under the service conditions formed by the updated air service node location, for the executable subtasks in the task information, first determine the target content source, and then determine the target computing node from the candidate computing nodes based on the total completion cost including the content acquisition cost from the target content source to the candidate computing node.

[0047] This step emphasizes the irreversible temporal order of "source first, then node" to avoid the redundant data transfer problem caused by selecting computing power first and then discovering that the content is unreachable, as seen in existing technologies. Specifically, for each executable subtask in the task information, its required content is determined. The set of candidate content sources is determined from the task requesting end, service nodes, air service nodes, ground edge nodes, and cloud servers. For each candidate content source n in the candidate content source set and each candidate computing node j in the candidate computing node set, determine the content acquisition cost for the candidate content source to provide the required content to the candidate computing node. As shown in the following formula: in, For content The data size (in bits) is obtained from the content metadata maintained by the content catalog server. The first term represents the transmission rate from the content source to the compute node, and the second term represents the transmission latency; the third term represents the cache miss penalty. The first term is the penalty coefficient; the second term is the penalty for link interruption risk. This is the penalty coefficient; The probability of link interruption from content source n to computing node j (for wireless links, the calculation method is the same). For wired links, take a fixed minimum value. ). and The method for unifying the dimensions converts latency (seconds) and probability (dimensionless) into a unified "cost unit" through a weighting factor. For example, setting 1 second of latency as equivalent to 0.1 unit of cost, then... and All are calibrated according to this standard. Specifically, This represents the equivalent latency penalty introduced when content misses the cache and needs to be retrieved from a remote node. It is set based on the difference in average retrieval latency between cache hits and misses in historical statistics. For example, if statistics show that the average additional latency for retrieving content from a remote node that has never missed the cache is 0.5 seconds, then... Setting it to 0.5 seconds automatically puts content sources that miss the cache at a disadvantage in the cost comparison, with the degree of disadvantage being consistent with the actual additional latency. This represents the equivalent latency penalty corresponding to the risk of link outage, set based on the average retransmission time after the outage. For example, if each link outage causes an average retransmission latency of 2 seconds, and the acceptable outage probability for the service is 1%, then... The time interval can be set to 200 seconds, so that a 1% increase in the probability of interruption corresponds to an equivalent latency cost of 2 seconds. Through the above calibration, the three components of the content acquisition cost correspond to the three dimensions of transmission latency, cache status, and link reliability, respectively, and are all converted to the time dimension, allowing for direct addition and comparison. This ensures that the selection of content sources simultaneously considers acquisition speed, cache hit rate, and link reliability. The candidate content source with the lowest content acquisition cost is determined as the target content source. .

[0048] After the target content source is determined, for each candidate computing node j in the candidate computing node set, the total completion cost of the candidate computing node is determined. The total cost of completion includes latency cost, energy cost, dependency penalty, and privacy penalty. Latency cost Energy consumption cost Dependence on punishment Privacy penalties Before weighted fusion, all data were normalized using historical statistics and mapped to... Interval. Specifically: Time delay normalization: Energy consumption normalization: ,in Battery capacity of task requester i Dependency penalty normalization: Privacy penalty normalization: (Privacy level is 5 at most) The normalized total completion cost is: in, to These are weighting coefficients, and their sum is 1. The weighting coefficients are set according to the needs of the business type: for latency-sensitive businesses such as vehicle-road cooperative early warning, the weighting coefficient is increased. The value of is chosen to make the total completion cost primarily driven by latency; for power-constrained task requesters, the value of is increased. The value of is chosen so that the total completion cost is primarily driven by energy consumption. In a specific example, it is set to... , , , This example assigns the highest weight to latency to meet the requirement of real-time task completion in highly dynamic scenarios; it assigns the second highest weight to dependency penalty to ensure the scheduler fully considers the propagation effect of predecessor delays in DAG tasks and reduces idle waiting for subtasks; and it assigns the lowest weight to privacy penalty because most subtasks have low data sensitivity, and only subtasks involving sensitive data will trigger significant privacy penalties. The calculation logic for each item is as follows: latency cost It covers the entire process of uploading, queuing, calculation, and content retrieval, as shown in the following formula: The above formula decomposes the completion latency of a subtask on a candidate computing node into the sum of four components: upload latency, queuing latency, computation latency, and content retrieval latency. Upload latency is determined by the ratio of the subtask's data volume to the upload transmission rate; queuing latency is the current queue waiting time of the candidate computing node; computation latency is determined by the ratio of the number of CPU cycles required by the subtask to the allocated computation frequency; and content retrieval latency, i.e., the content retrieval cost mentioned above, represents the time it takes for the required content to travel from the target content source to the candidate computing node. Including all four components in the total latency is a conservative estimate obtained by processing content retrieval and data upload sequentially. In actual execution, content retrieval and data upload can be performed in parallel, so the actual completion latency is usually no greater than this estimate. The significance of using a conservative estimate is that computing nodes selected based on this estimate have a latency margin in actual execution, resulting in higher feasibility of the scheduling outcome and avoiding subtask timeouts caused by optimistic estimations.

[0049] Energy consumption cost The energy consumption for both transmission and computing is shown in the following formula: In the above formula, the first term represents the transmission energy consumption generated by the task requesting end uploading subtask data, which is equal to the product of the transmission power and the upload latency. The second term represents the computational energy consumption generated by the candidate computing node executing the subtask. The computational frequency in the computational energy consumption is expressed in a square form, derived from the dynamic power consumption model of CMOS circuits: the processor's dynamic power consumption is proportional to the cube of the computational frequency, while the computational latency is equal to the ratio of the required CPU cycles to the computational frequency. Multiplying the power consumption by the computational latency yields the relationship that computational energy consumption is proportional to the square of the computational frequency, i.e., the form of the second term in the above formula. This relationship reveals the characteristic of energy consumption increasing non-linearly with computational frequency: while increasing the allocated computational frequency can linearly shorten the computational latency, it will increase computational energy consumption in a square relationship. Therefore, the allocation of computational frequency needs to balance latency and energy consumption, rather than simply pursuing the maximum value. The calculated energy consumption coefficient for node j is determined by the effective capacitance and supply voltage of the node hardware, and can be obtained by consulting the hardware manual or by actual measurement and calibration on a typical load. The computation frequency allocated to node j for this subtask is determined by a proportional allocation method: the allocable computation frequency of node j is allocated according to the proportion of computation of each subtask waiting to be executed on node j, and the sum of the computation frequencies allocated to each subtask is guaranteed not to exceed the total available computation frequency of node j, so as to fairly and feasiblely allocate computing resources among multiple subtasks.

[0050] Dependency penalty The pre- and post-dependencies used to handle DAG tasks are shown in the following equation: in, This is the time when the precursor sub-task is completed. This represents the scheduled start time of the current subtask. For example, if the predecessor task is delayed by 2 seconds due to network fluctuations, and the current subtask was originally scheduled to start in 1 second, a 1-second dependency penalty will occur. This ensures that the scheduler can perceive the propagation effect of the predecessor delay, avoiding blind scheduling that leads to idle waiting.

[0051] Privacy penalties It only applies to non-local nodes, as shown in the following formula: in, Privacy levels range from 1 to 5, with level 5 being the highest sensitivity level. L represents the type identifier of the local processing node at the task request end. Privacy levels are predefined based on business type: the system maintains a mapping relationship between business types and privacy levels, and the privacy level of a subtask is determined by the type of data it processes. For example, a subtask processing public traffic media streams is set to level 1, a subtask processing vehicle location trajectories is set to level 3, and a subtask processing payment authentication or biometric data is set to level 5. The rationale for this setting is that the degree of harm from the leakage of different types of data varies. Quantifying privacy requirements into comparable numerical levels allows privacy factors to be directly incorporated into cost calculations, and can be uniformly weighed against indicators such as latency and energy consumption. This ensures that sensitive data is processed locally when external nodes cannot provide sufficient benefits. This is an indicator function. When The penalty is 0 when executed locally; when offloaded to an external node, the higher the privacy level, the greater the penalty. This forces sensitive data to be processed locally first, unless the latency / energy advantage provided by the external node is sufficient to offset the risk of privacy leakage. Ultimately, the candidate computing node with the lowest total completion cost is determined as the target computing node.

[0052] Step 5: Perform end-to-end reliability verification on the execution paths formed from the target content source to the target computing node and from the task request end to the target computing node, and lock communication resources and computing resources after the execution path meets the reliability conditions.

[0053] To ensure the theoretically optimal path is usable in a real physical environment, this step introduces hard reliability constraints. The link between the task requester and the target computing node is defined as the uplink segment, and the link between the target content source and the target computing node is defined as the content transmission link segment. The interruption probabilities of the two links are obtained, and the end-to-end interruption probability is determined based on these probabilities. As shown in the following formula: in, This indicates the number of link segments included in the execution path, which in this embodiment includes two segments: uplink link segment and content transmission link segment. Indicates the first The probability of link segment interruption is calculated using the same method as the link interruption probability mentioned above, based on the signal-to-noise ratio and interruption threshold of that link segment. The product form is based on the reliability model of a serial system: the execution path consists of multiple hops connected in series, and the interruption of any link will lead to overall failure. End-to-end reliability is only achieved when all links are reliable. The system then determines whether the end-to-end interruption probability does not exceed a preset reliability threshold. Preset reliability threshold Set according to the service level requirements of the business, for example, take 0.01. For security-related businesses that require a reliability of no less than 99%, The threshold is set to 0.01; for non-security services with high latency tolerance, it can be relaxed to 0.05. A stricter threshold setting results in higher execution path reliability, but also a higher probability of triggering a reselection of the target computing node, increasing scheduling overhead. Therefore, the threshold needs to be set according to the service's tolerance for reliability and scheduling overhead. If the threshold is not exceeded, the target computing node is retained, and its computing frequency resources and link bandwidth resources are locked. The locking operation specifically includes computing frequency resource locking and link bandwidth resource locking. Computing frequency resource locking requires that the sum of the computing frequencies allocated to each subtask by the node does not exceed the total available computing frequencies of that node, as shown in the following formula: in, This represents the total available computational frequency of node j. This indicates the subtask assigned by node j to task requester i. The computation frequency is calculated, and the summation is performed on all subtasks targeting node j. This formula ensures that the computation frequency resource is locked within the computing power capacity of the node, preventing over-selling of computing power and resulting in the actual computation speed of each subtask being lower than the estimated value. Link bandwidth resource locking adopts a proportional allocation method. Let the total bandwidth of node j be... Assign to subtask The bandwidth ratio is If the sum of the bandwidth proportions occupied by each subtask does not exceed 1, as shown in the following formula: in, This represents the total allocatable bandwidth of node j. Subtasks The proportion of the total bandwidth in node j. The above formula ensures that the bandwidth allocation does not exceed the physically available bandwidth in the form of a proportional constraint. After the bandwidth proportion is locked, the actual transmission rate from the task requesting end to the target computing node is determined according to the following formula: in, This represents the channel gain from the task requester to node j. The logarithmic term represents the achievable capacity per unit bandwidth given by Shannon's formula. Multiplying this by the bandwidth ratio and the total bandwidth yields the actual transmission rate occupied by the subtask. This formula converts the locked bandwidth ratio into an executable transmission rate, ensuring that the latency estimate for subsequent cycles matches the actual transmission capacity. The rationale for using the locking mechanism is that if resources are not locked in time after the target computing node is determined, other scheduling requests may occupy the same portion of resources, leading to resource shortages during the actual execution of the subtask and rendering previously determined low-cost paths unexecutable. Locking ensures the dedicated use of resources during the execution of the subtask, making the cost estimate consistent with the actual execution result and improving the determinism of the scheduling result. If the threshold is exceeded, the target computing node is removed from the candidate computing node set, and the process returns to step 4 to reselect a suboptimal node until a path meeting the conditions is found or the number of retries reaches the upper limit (e.g., 3 times). This mechanism eliminates paths that are "theoretically low-cost but prone to disconnection in practice," ensuring the determinism of task execution.

[0054] Furthermore, after locking the computing frequency resources and link bandwidth resources, the process also includes: recording the difference between scheduling characteristics and actual costs, generating local correction parameters, and updating the estimation model for the next cycle after asynchronous uploading and aggregation with the cloud side.

[0055] To address the discrepancy between model predictions and actual conditions, this step establishes a parameter-level asynchronous correction closed loop. Specifically, after locking resources and executing subtasks by the task requester, local samples are formed on a per-subtask basis. Each local sample consists of the difference between scheduling features and actual costs, where the scheduling features are organized as feature vectors. Defined by the following formula: The feature vector comprises ten dimensions: subtask data volume, subtask computational load, subtask privacy level, task requester speed, task requester heading, overall task risk value, remaining coverage time, link interruption probability, target computing node queue waiting time, and content cache hit status. These features were selected because they cover three categories of factors that determine the deviation between predicted and actual costs. Data volume, computational load, and privacy level are task attributes that determine the baseline cost; speed, heading, task risk value, remaining coverage time, and interruption probability are movement and link states that determine transmission cost fluctuations; and queue waiting time and cache hit status are resource states that determine computation and content retrieval cost fluctuations. Furthermore, each dimension of the feature vector can be obtained locally by the task requester without needing to acquire raw data from other nodes. Actual cost difference This represents the difference between the actual execution cost of a subtask and the predicted cost during scheduling. The predicted cost is the total completion cost calculated during scheduling, while the actual execution cost is obtained by weighted summation of measured completion latency, measured energy consumption, and other metrics after the subtask is executed. A positive difference indicates that the actual cost is higher than the prediction, suggesting an optimistic cost estimate for this type of scheduling scenario; a negative difference indicates that the actual cost is lower than the prediction, suggesting a conservative cost estimate. Local correction parameters are generated based on these data. Specifically, this includes: generating local correction parameters from a lightweight correction model deployed locally on the task requesting end. The lightweight correction model is a feedforward neural network, consisting of an input layer, a feature normalization layer, a first fully connected layer, a second fully connected layer, and an output layer arranged sequentially.

[0056] The input layer receives the difference between the feature vector and the actual cost from the local samples, and has eleven input nodes. The role of the input layer is to receive and transmit the original scheduling samples; it does not perform calculations itself, but provides a unified data entry point for subsequent layers.

[0057] The function of the feature normalization layer is to standardize the features in each dimension. This involves subtracting the mean of each feature from its standard deviation and then dividing by the mean of that dimension, thus eliminating dimensional differences between features. For example, data volumes can range from thousands of bytes to tens of megabytes, while the probability of interruption falls between 0 and 1. Without normalization, features with large numerical ranges will dominate the gradient update direction, causing training to favor those features and hindering convergence. The mean and standard deviation used for normalization are obtained from local historical samples and are updated periodically as samples accumulate, ensuring that the normalization benchmark always matches the current operating environment.

[0058] The first fully connected layer maps the normalized eleven-dimensional features to thirty-two hidden layer features, employing the ReLU activation function. The role of this first fully connected layer is to perform non-linear combinations of the input features, extracting low-order interactions between features, such as the joint effect of speed and remaining coverage time on cost bias. The ReLU activation function outputs zero for negative inputs and remains unchanged for positive inputs. Its computation is simple and it does not suffer from gradient saturation in the positive interval, making it suitable for deployment in computationally limited vehicle terminals. The reason for choosing thirty-two hidden layer neurons is that with eleven input features, the hidden layer width expands to approximately three times the input dimension, providing ample feature combination space while keeping the number of parameters in the hundreds, meeting lightweight requirements.

[0059] The second fully connected layer maps the 32-dimensional hidden layer features to 16-dimensional hidden layer features, also using the ReLU activation function. The role of the second fully connected layer is to perform a secondary abstraction of the low-order features, forming a high-order representation related to the cost bias pattern. The dimension is gradually compressed from 32 to 16, which can compress redundant information and suppress overfitting. Since the number of local samples on the vehicle side is limited, a network that is too deep or too wide is prone to remembering noise from the samples and losing its generalization ability; therefore, a two-layer structure with moderate width is adopted.

[0060] The output layer maps the 16-dimensional hidden layer features to a 4-dimensional output using a linear activation function. The four output nodes correspond to the cost correction amounts for four types of candidate execution nodes: local processing nodes, service nodes, over-the-air service nodes, and ground edge nodes. The output of the output layer represents the local correction parameters. The reason for using a linear activation function in the output layer is that the correction value is a real number that can be positive or negative. Positive values ​​are used to adjust the cost estimate upwards, and negative values ​​are used to adjust the cost estimate downwards. It should not be compressed into a fixed interval by a nonlinear activation function.

[0061] The training process of the lightweight correction model is as follows: Scheduling features from local samples are used as input, the difference between actual costs in local samples is used as the supervision target, and the mean squared error between the network output and the actual cost difference is used as the loss function. The Adam optimizer is used to iteratively update the network parameters through backpropagation. The learning rate is set to 0.001, based on the fact that this value is a common order of magnitude for the Adam optimizer, ensuring convergence speed while minimizing gradient oscillations. Each training iteration uses recently accumulated local samples, with a batch size of sixteen. If the loss function no longer decreases significantly after several consecutive rounds, training is terminated early to save computational resources for the vehicle terminal. During inference, the scheduling features of the current subtask are input into the trained lightweight correction model. The four correction values ​​output are then superimposed on the estimated costs of the four types of candidate execution nodes. This means that the minimum remaining processing time estimate and the total completion cost estimate in the next control cycle are both corrected using these correction values, making the estimated cost closer to the actual cost. The reason for using a neural network structure instead of nonlinear regression is that the mapping relationship between cost bias and features such as mobility and resource status is nonlinear. For example, when the speed exceeds a certain threshold, the impact of Doppler frequency shift on the link rate exhibits abrupt changes, making it difficult for a linear model to fit. A two-layer fully connected network can approximate this nonlinear mapping with sufficient accuracy and has a small number of parameters, making it suitable for vehicle-side deployment and parameter-level uploading.

[0062] To avoid signaling storms caused by simultaneous uploads from multiple nodes, a preset random backoff time is allowed after the local parameter correction is generated. Then upload it to the cloud. Follow the interval The uniform distribution of data on the network effectively prevents collisions by dispersing upload requests that might be triggered simultaneously after each control cycle, thus avoiding signaling storms and uplink congestion caused by multiple nodes uploading at the same time. The asynchronous upload time of task request terminal i... Determined by the following formula: in, This indicates the time when task requesting terminal i completes local training. The backoff time is random. Maximum backoff window. The settings are based on the cloud's upload processing capacity and the number of task request clients in the system. For example, if the cloud can process K upload requests per second, and there are N task request clients in the system, then... Setting it to N / K seconds ensures that upload requests generated within the backoff window can be fully processed by the cloud side on average, preventing upload requests from accumulating in the cloud side queue. After receiving the parameters, the cloud side adjusts the parameters based on the number of samples from the task requesting end that uploaded the locally corrected parameters. Perform weighted aggregation to obtain the global correction parameters. As shown in the following formula: in, This refers to the set of task requesters that have completed the generation of locally corrected parameters and are ready to upload them within the current control period. (Sample size weighting) The significance lies in giving nodes with higher data quality greater say: the more samples, the more stable the statistical regularity, and the more reliable the parameter correction. The aggregated global parameters are distributed to each node to update the minimum remaining processing time estimate and total completion cost estimate in the next control cycle, enabling the system to have continuous learning capabilities.

[0063] Furthermore, after updating the minimum remaining processing time estimate and total completion cost estimate in the next control cycle using global correction parameters, the process also includes: rolling out the scheduling process based on the updated parameters until the termination condition is met.

[0064] In the next control cycle, based on the updated air service node location, link status, queue status, cache status, and global correction parameters, steps 1 to 5 are re-executed to obtain the comprehensive optimization target. As shown in the following formula: In the above formula, the summation is performed on all subtasks of all requesting tasks. Subtasks The completion delay Subtasks Total energy consumption Subtasks End-to-end interruption probability of the execution path Subtasks The privacy penalties are categorized into four indicators, corresponding to four optimization dimensions: timeliness, energy efficiency, reliability, and privacy. to These are weighting coefficients that sum to 1, set according to the system optimization objective. In a specific example, they are set as follows: , , , This example assigns the highest weight to latency because task timeout is the most direct manifestation of performance degradation in highly dynamic scenarios; it assigns the second highest weight to interruption probability because air-to-ground link interruptions occur frequently and have a significant impact; energy consumption and privacy penalties are assigned lower weights. The setting method is similar to the calibration method for risk fusion weights: first, regression analysis is performed on the correlation between each indicator and task failure rate based on historical data, and then fine-tuning is done in conjunction with business priorities to ensure that the comprehensive optimization target truly reflects the core requirements of the current scenario. It then determines whether all executable subtasks have been completed. If not, and the decrease in the comprehensive optimization target relative to the previous control cycle is greater than or equal to a preset threshold... If all subtasks are completed, or the decrease is less than 1%, then the next control cycle will continue to be executed; If the condition is met, task scheduling ends. Convergence threshold. The setting is based on the system jitter tolerance, for example, set to 1% of the target value. The reason for this setting is that, in the absence of substantial optimization benefits, the overall optimization target will still experience slight jitter between adjacent control cycles due to measurement noise and channel fluctuations. Using the jitter amplitude as a lower bound for effective improvement, it can be determined that continuing iteration will not bring substantial improvement when the decrease amplitude is lower than this lower bound, thereby terminating the scheduling in time and avoiding ineffective iteration that consumes computing and communication resources.

[0065] It should be noted that the preset parameters involved in the above steps are not fixed values. The zero-minimum constant is set according to machine accuracy, for example, 10 to the power of -6 to 10 to the power of -9; the prediction time window is determined based on the ratio of the maximum speed of the task requesting end to the coverage radius, for example, 0.5 seconds to 2.0 seconds; the risk fusion weight satisfies a sum of 1, determined by a combination of historical data regression and expert experience fine-tuning; the potential field anti-singularity term is taken as 0.5% to 2% of the coverage radius, determined through numerical stability testing; the potential field forward step length is taken as 30% to 60% of the maximum flight distance of the air service node in each control cycle, determined based on the air service node's maneuverability and control cycle matching; the node repulsion coefficient is taken as 0.1 to 1.0, determined through simulation debugging to balance aggregation risk and coverage efficiency; the content acquisition penalty coefficient is taken as 0.1 to 10.0, determined through dimensional transformation and Pareto front search; the privacy level is taken as 1 to 5, predefined according to the service type; the reliability threshold is taken as 0.001 to 0.05, determined according to the service level requirements of the business; the convergence threshold is taken as 0.1% to 5% of the target value, determined based on system jitter tolerance testing.

[0066] Please see Figure 2 The present invention further provides a UAV mission unloading and trajectory planning system, the system being used to implement the above-mentioned UAV mission unloading and trajectory planning method, comprising: The data acquisition module is used to acquire the status information and task information of the task requesting end, predict the predicted location of the task requesting end based on the status information, and determine the remaining coverage time of the task requesting end relative to the current service air node. The risk assessment module is used to generate task risk values ​​based on the predicted location, remaining coverage time, and task information corresponding to the task execution risk. The location prediction module is used to generate a spatial risk potential field based on the mission risk value and the predicted location, and to control the next moment's location of the air service node according to the spatial risk potential field. The target determination module is used to determine the target content source for executable subtasks in the task information under the service conditions formed by the updated air service node locations, and then determine the target computing node from the candidate computing nodes based on the total completion cost including the content acquisition cost from the target content source to the candidate computing node. The resource verification module is used to perform end-to-end reliability verification on the execution path formed from the target content source to the target computing node and from the task request end to the target computing node, and lock communication resources and computing resources after the execution path meets the reliability conditions.

[0067] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0068] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0069] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0070] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that cannot be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for unloading and trajectory planning of unmanned aerial vehicle (UAV) missions, characterized in that, The specific steps include: Obtain the status information and task information of the task requester, predict the predicted location of the task requester based on the status information, and determine the remaining coverage time of the task requester relative to the current service air node. Based on the predicted location, remaining coverage time, and task information, generate a task risk value corresponding to the task execution risk. A spatial risk potential field is generated based on the mission risk value and the predicted location, and the next moment's location of the air service node is controlled according to the spatial risk potential field. Under the service conditions formed by the updated air service node location, for the executable subtasks in the task information, first determine the target content source, and then determine the target computing node from the candidate computing nodes based on the total completion cost including the content acquisition cost from the target content source to the candidate computing node. Perform end-to-end reliability verification on the execution path formed from the target content source to the target computing node and from the task request end to the target computing node, and lock communication resources and computing resources after the execution path meets the reliability conditions.

2. The method for unloading and trajectory planning of a UAV mission according to claim 1, characterized in that: Based on the predicted location, remaining coverage time, and task information, corresponding to the task execution risk, a task risk value is generated, including: Based on the completion status of subtasks in the task information, determine the set of currently executable subtasks from the task information; For each currently executable subtask in the current set of executable subtasks, estimate the minimum remaining processing time; Based on the minimum remaining processing time and the latest completion time of the currently executable subtasks, generate task deadline risks; Coverage handover risk is generated based on the remaining coverage time, where the shorter the remaining coverage time, the higher the coverage handover risk. Generate link interruption risk based on the link quality between the task requester and the current service air node; Queue waiting risk is generated based on the queue waiting time of candidate computing nodes; Generate cache mismatch risk based on the cache hit status of the content required by the currently executable subtask on the candidate content source; The risk values ​​of task deadline, coverage switching, link interruption, queue waiting, and cache mismatch are weighted and fused to obtain the task risk value.

3. The method for unloading and trajectory planning of unmanned aerial vehicle (UAV) missions according to claim 2, characterized in that: Based on the minimum remaining processing time and the latest completion deadline of the currently executable subtasks, generate task deadline risks, including: Local processing nodes, service nodes, air service nodes, and ground edge nodes are selected as candidate execution nodes; For each candidate execution node, the estimated processing time of the current executable subtask on the candidate execution node is determined based on the available transmission rate from the task requester to the candidate execution node, the amount of data of the current executable subtask, the allocable computing frequency of the candidate execution node, the amount of computing required by the current executable subtask, and the queue waiting time of the candidate execution node. Select the minimum value from the estimated processing time corresponding to the candidate execution nodes as the minimum remaining processing time; Calculate the ratio between the minimum remaining processing time and the remaining time limit of the currently executable subtask, where the remaining time limit is the difference between the latest completion time limit and the current time. The largest ratio in the current set of executable subtasks is determined as the task deadline risk; Based on the remaining coverage time, coverage handover risks are generated, including: Compare the remaining coverage time with the preset prediction time window; Increase the risk of coverage handover when the remaining coverage time is less than the preset prediction time window; Reduce the risk of coverage handover when the remaining coverage time is greater than or equal to the preset prediction time window; Based on the link quality between the task requester and the current service air node, link interruption risks are generated, including: Obtain the signal-to-noise ratio or average signal-to-noise ratio between the task requester and the current service's air service node; Increase the risk of link interruption when the signal-to-noise ratio or average signal-to-noise ratio is lower than or equal to a preset interruption threshold; Reduce the risk of link interruption when the signal-to-noise ratio or average signal-to-noise ratio is higher than the preset interruption threshold; Queue waiting risk is generated based on the queue waiting time of candidate computing nodes, including: Get the current queue waiting time of the candidate computing node; Increase the risk of queue waiting when the queue waiting time increases relative to the previous detection time; Reduce queue waiting risk when the queue waiting time decreases relative to the previous detection time; Based on the cache hit status of the content required by the currently executable subtask on the candidate content source, a cache mismatch risk is generated, including: Determine whether the content required by the currently executable subtask is cached in at least one candidate content source; Increase the risk of cache mismatch if the required content is not cached in any of the candidate content sources; Reduce the risk of cache mismatch when the required content is cached in at least one candidate content source.

4. The method for unloading and trajectory planning of unmanned aerial vehicle (UAV) missions according to claim 1, characterized in that: A spatial risk potential field is generated based on the mission risk value and predicted location, and the next-time location of the airborne service node is controlled according to the spatial risk potential field, including: The predicted location of the task requester is used as the risk center; Multiple spatial location points are set within the service area plane; Based on the task risk value and the distance between the spatial location point and the predicted location, the risk potential value of each spatial location point is determined. The higher the task risk value, the greater the risk potential value; the closer the spatial location point is to the predicted location, the greater the risk potential value. The direction of movement of the air service node is determined based on the increasing risk potential at its current location; Update the planar position of the air service node according to the direction of movement to obtain the next position of the air service node.

5. The method for unloading and trajectory planning of a UAV mission according to claim 4, characterized in that: Update the planar position of the air service node according to the direction of movement to obtain the next position of the air service node, including: Obtain the gradient direction of the risk potential field at the current location of the air service node; Move along the gradient direction of the risk potential field by a preset step size to obtain the initial position for the next time step. Obtain the location of other air service nodes; Based on the distance between the location of other air service nodes and the current location of the air service node, a repulsion displacement is generated. The smaller the distance between the current location and other air service nodes, the larger the repulsion displacement. Based on the initial next-time position and the repulsion displacement, candidate next-time positions are generated; Determine whether the displacement length between the candidate next position and the current position of the air service node exceeds the preset maximum flight distance; If the displacement length exceeds the preset maximum flight distance, the displacement vector of the candidate next moment position relative to the current position is scaled proportionally according to the ratio of the preset maximum flight distance to the displacement length to obtain the next moment position. If the displacement length does not exceed the preset maximum flight distance, the candidate position for the next moment is determined as the position for the next moment.

6. The method for unloading and trajectory planning of a UAV mission according to claim 1, characterized in that: For executable subtasks in the task information, first determine the target content source, then, based on the total completion cost including the content acquisition cost from the target content source to the candidate computing nodes, determine the target computing node from the candidate computing nodes, including: For each executable subtask in the task information, determine the required content for the executable subtask; The candidate content source set is determined from the task request end, service node, air service node, ground edge node and cloud, and the candidate computing node set is determined from the service node, air service node, ground edge node and cloud; For each candidate content source in the candidate content source set and each candidate computing node in the candidate computing node set, determine the content acquisition cost for the candidate content source to provide the required content to the candidate computing node; The candidate content source with the lowest content acquisition cost is identified as the target content source; After the target content source is determined, for each candidate computing node in the candidate computing node set, the total completion cost of the candidate computing node is determined. The total completion cost includes the transmission cost of the task requester uploading subtask data to the candidate computing node, the queuing cost of the candidate computing node, the computing cost of the candidate computing node, and the content acquisition cost of the target content source providing the required content to the candidate computing node. The candidate computing node with the lowest total completion cost is selected as the target computing node.

7. The method for unloading and trajectory planning of a UAV mission according to claim 6, characterized in that: The total cost of completion also includes energy costs, dependency penalties, and privacy penalties; Determine the total completion cost of candidate computing nodes, including: The latency cost is determined based on the data upload time from the task requester to the candidate computing node, the queue waiting time of the candidate computing node, the computing time of the candidate computing node for the executable subtask, and the content acquisition time of the target content source to transmit the required content to the candidate computing node. The energy cost is determined based on the transmission energy consumption generated by the task requester uploading executable subtask data and the computing energy consumption generated by the candidate computing nodes executing executable subtasks. Get the completion time of the predecessor subtask of the executable subtask; When the completion time of the predecessor subtask is later than the planned start time of the executable subtask on the candidate computing node, the dependency penalty is determined based on the difference between the completion time and the planned start time of the predecessor subtask. The larger the difference, the larger the dependency penalty. When the candidate computing node is not local to the task requester, the privacy penalty is determined based on the privacy level of the executable subtask; the higher the privacy level, the greater the privacy penalty. The total completion cost is determined by a weighted sum of latency cost, energy cost, dependency penalty, and privacy penalty.

8. The method for unloading and trajectory planning of a UAV mission according to claim 6, characterized in that: Perform end-to-end reliability verification on the execution paths formed from the target content source to the target computing node and from the task requester to the target computing node, and lock communication and computing resources after the execution path meets the reliability conditions, including: The link between the task requester and the target computing node is defined as the uplink segment. The link between the target content source and the target computing node is defined as the content transmission link segment; Obtain the probability of uplink segment interruption and the probability of content transmission segment interruption; The end-to-end interruption probability is determined based on the interruption probability of the uplink segment and the interruption probability of the content transmission segment. The higher the interruption probability of the uplink segment, the higher the interruption probability of the content transmission segment, and the higher the end-to-end interruption probability. Determine whether the end-to-end interruption probability does not exceed a preset reliability threshold; If the end-to-end interruption probability does not exceed the preset reliability threshold, the target computing node is retained, and the computing frequency resources and link bandwidth resources of the target computing node are locked. If the end-to-end interruption probability exceeds a preset reliability threshold, the target computing node is removed from the candidate computing node set, and the process of determining the target content source and target computing node is returned until the end-to-end interruption probability corresponding to the re-determined execution path does not exceed the preset reliability threshold.

9. The method for unloading and trajectory planning of a UAV mission according to claim 8, characterized in that: After locking the computing frequency resources and link bandwidth resources of the target computing node, the following is also included: Record the difference between scheduling features and actual costs. Scheduling features include at least one of the following: data volume of executable subtasks, computational volume, privacy level, speed of task requesting end, heading, task risk value, remaining coverage time, link interruption probability, queue waiting time, and content cache hit status. The actual cost difference is the difference between the actual execution cost and the predicted cost. Local correction parameters are generated based on the difference between scheduling characteristics and actual costs. After the local correction parameters are generated, a preset random backoff time is allowed before uploading the local correction parameters to the cloud. Receive global correction parameters from the cloud side. The global correction parameters are obtained by the cloud side by weighting and aggregating the local correction parameters based on the number of samples from the task requesting end that uploaded the local correction parameters. Update the minimum remaining processing time estimate and total completion cost estimate in the next control cycle using global correction parameters; After updating the minimum remaining processing time estimate and total completion cost estimate for the next control cycle using the global correction parameters, the following steps are also included: In the next control cycle, based on the updated air service node location, link status, queue status, cache status and global correction parameters, the process of predicting location, task risk value, spatial risk potential field, target content source, target computing node, end-to-end reliability verification and locking of communication resources and computing resources will be re-executed. Obtain the comprehensive optimization target, which is determined based on subtask completion latency, energy consumption, end-to-end interruption probability, and privacy penalty. Determine whether all executable subtasks have been completed; If all executable subtasks are not completed and the overall optimization target decreases by a factor greater than or equal to a preset threshold relative to the previous control cycle, the next control cycle will continue to be executed. Task scheduling ends when all executable subtasks are completed, or when the overall optimization target decreases less than a preset threshold relative to the previous control cycle.

10. A UAV mission unloading and trajectory planning system, characterized in that: The system is used to implement the UAV mission unloading and trajectory planning method according to any one of claims 1-9, including: The data acquisition module is used to acquire the status information and task information of the task requesting end, predict the predicted location of the task requesting end based on the status information, and determine the remaining coverage time of the task requesting end relative to the current service air node. The risk assessment module is used to generate task risk values ​​based on the predicted location, remaining coverage time, and task information corresponding to the task execution risk. The location prediction module is used to generate a spatial risk potential field based on the mission risk value and the predicted location, and to control the next moment's location of the air service node according to the spatial risk potential field. The target determination module is used to determine the target content source for executable subtasks in the task information under the service conditions formed by the updated air service node locations, and then determine the target computing node from the candidate computing nodes based on the total completion cost including the content acquisition cost from the target content source to the candidate computing node. The resource verification module is used to perform end-to-end reliability verification on the execution path formed from the target content source to the target computing node and from the task request end to the target computing node, and lock communication resources and computing resources after the execution path meets the reliability conditions.

Images