Cache-assisted unmanned aerial vehicle edge computing unloading decision and resource allocation optimization method

Abstract

The invention discloses a cache-assisted unmanned aerial vehicle edge computing unloading decision and resource allocation optimization method, and belongs to the technical field of wireless communication. Comprising the following steps: deploying a plurality of unmanned aerial vehicles as an edge server, determining an initial deployment position of each unmanned aerial vehicle by adopting a clustering method based on a user position, executing a local search strategy around the initial deployment position, and determining a final hovering position and an initial cache of each unmanned aerial vehicle by taking maximization of a user coverage rate as a target; based on the final hovering position and the initial cache of the unmanned aerial vehicle, a joint optimization problem with the aim of minimizing the average total cost of the multi-unmanned aerial vehicle edge computing system is constructed, and the joint optimization problem is decomposed into a discrete unloading decision sub-problem and a continuous computing resource allocation sub-problem; and solving the unloading decision sub-problem and the computing resource allocation sub-problem by adopting an alternating optimization strategy based on a genetic method and a particle swarm method, and obtaining a joint optimal solution through iteration.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a cache-assisted method for unloading decision-making and resource allocation optimization in edge computing for unmanned aerial vehicles. Background Technology

[0002] With the rapid development of IoT applications, the number of smart terminal devices has surged, generating a massive number of computationally intensive and latency-sensitive tasks. However, the computing power, storage capacity, and battery life of these terminal devices are limited, making it difficult for them to independently handle heavy computational loads. Mobile edge computing, as an emerging computing paradigm, effectively reduces task processing latency and core network bandwidth pressure by pushing computing resources down to the network edge, closer to the user. Drones, with their high mobility, flexible deployment, and line-of-sight communication links, are becoming important nodes in building integrated air-ground mobile edge computing networks. Using drones as aerial mobile base stations and edge servers can provide dynamic, on-demand computing services to areas where ground infrastructure coverage is insufficient.

[0003] While existing research has extensively explored UAV-assisted mobile edge computing, several shortcomings remain. For example, some studies employ static computing resource allocation strategies, meaning resources remain unchanged after allocation, failing to dynamically adjust based on task completion and the addition of new tasks, leading to low resource utilization. Furthermore, some studies neglect the reality that tasks may span multiple time slots, simplifying the problem model. Simultaneously, how to jointly optimize UAV deployment locations, task offloading decisions, and dynamic resource allocation to minimize system costs remains a challenging problem.

[0004] Therefore, there is an urgent need for a method that can adapt to the dynamic arrival characteristics of missions, jointly optimize UAV deployment, offloading decisions, and dynamic resource allocation, in order to improve the performance of the entire UAV edge computing system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a cache-assisted method for unloading decision-making and resource allocation optimization in UAV edge computing, thereby solving the problems in existing technologies.

[0006] The objective of this invention can be achieved through the following technical solutions: A cache-assisted method for drone edge computing offloading decision-making and resource allocation optimization, wherein the multi-drone edge computing system includes: a base station, multiple drones, and multiple ground users, and the method includes the following steps: Multiple drones are deployed as edge servers. A clustering method based on user location is used to determine the initial deployment location of each drone. A local search strategy is then executed around the initial deployment location to maximize user coverage and determine the final hovering location and initial cache of each drone. Based on the final hovering position and initial cache of the UAV, a joint optimization problem is constructed with the goal of minimizing the average total cost of the multi-UAV edge computing system. The joint optimization problem is decomposed into a discrete offloading decision subproblem and a continuous computing resource allocation subproblem. An alternating optimization strategy based on genetic methods and particle swarm optimization is used to solve the unloading decision subproblem and the computational resource allocation subproblem, and the joint optimal solution is obtained through iteration.

[0007] Furthermore, the process of determining the final hovering position of each drone includes: All users are divided into groups using the K-Means algorithm. non-overlapping clusters This minimizes the sum of squared Euclidean distances between users within a cluster, for each cluster. Calculate its geometric centroid and use it as the drone initial position ; drones The user coverage area is divided into a grid. All candidate locations in the grid are traversed, the number of users covered at each location is calculated, and the location with the highest coverage is selected as the final hovering location for the drone. Tasks generated by users not covered by drones will be discarded.

[0008] Furthermore, the process of determining the initial cache is as follows: Each drone Equipped with a capacity of The drone pre-stores a portion of popular service programs in its cache space using a binary variable. Indicates drone Is the service cached? ,in Indicates service Cache on drones Up, or vice versa Cache decisions must satisfy the drone's capacity constraints, expressed as: in, It is a service The size of the drone set is K, where K represents the drone set.

[0009] Furthermore, the joint optimization problem is: in, These represent the drone caching decision, user offloading decision, and computing resource allocation sets, respectively. express The set of tasks at any given moment; For the total number of tasks, For a set of tasks, For caching decisions, It is a service The size of K represents the drone set. For unloading decision set, For drones computing power Indicates time slot drones For the task Allocated computing resources; The total system cost, As a weighting factor for user fees, For system task completion time, For user fees; For the latency of user uploads to the drone, For the computational latency of the drone, For the mission from drone Transmit to drone The time delay, For the service program from Transmit to The time delay, This refers to the latency of downloading service programs from the base station.

[0010] Furthermore, the unloading decision subproblem is: in, These represent the user uninstallation decision set, For the total number of tasks, For a set of tasks, The total system cost, express The set of tasks at any given moment; For the unloading decision set.

[0011] Furthermore, the computational resource allocation subproblem is as follows: in, Represents the set of computing resource allocations. For the total number of tasks, For a set of tasks, The total system cost, express Drones The set of tasks executed on top For drones computing power Indicates time slot drones For the task Allocated computing resources.

[0012] Furthermore, the process of solving the unloading decision subproblem and the computational resource allocation subproblem using an alternating optimization strategy includes: S31, Under a fixed weight of computing resource allocation, the genetic method is used to solve the unloading decision subproblem to obtain the optimal unloading decision combination; S32, fix the obtained unloading decision, use the particle swarm optimization method to solve the computational resource allocation subproblem, and obtain the optimal computational resource allocation weight; S33, iteratively execute until the total system cost converges, and obtain the globally optimal joint scheme of unloading decision and computing resource allocation.

[0013] Furthermore, both the genetic method and the particle swarm optimization method employ an integrated event-driven simulator as the fitness function; the fitness function expression is: in, For the fitness function, Represents the set of old tasks. Represents a new set of tasks. For the task The computational latency required, For each new task Total communication latency, As a weighting factor for user fees, For user fees, This is the final fitness value of the joint scheme; For the task The remaining computational load, Indicates drone The task queue on the screen, Indicates drone Computing resources For the task Resource weight, Indicates task The allocated computing resources; These are the transmission delays of D2U, U2U, and B2U links obtained based on the hybrid NOMA rate model.

[0014] A cache-assisted drone edge computing offloading decision and resource allocation optimization device performs the above method, including: Location confirmation module: Deploy multiple drones as edge servers, use a clustering method based on user location to determine the initial deployment location of each drone, and execute a local search strategy around the initial deployment location to maximize user coverage, and determine the final hovering location and initial cache of each drone. Problem Construction and Decomposition Module: Based on the final hovering position and initial cache of the UAV, a joint optimization problem is constructed with the goal of minimizing the average total cost of the multi-UAV edge computing system, and the joint optimization problem is decomposed into a discrete offloading decision subproblem and a continuous computing resource allocation subproblem; Problem-solving module: It uses an alternating optimization strategy based on genetic methods and particle swarm optimization to solve the unloading decision subproblem and the computational resource allocation subproblem, and obtains the joint optimal solution through iteration.

[0015] A computer storage medium storing a readable program that, when executed, instructs a computing device to perform a cache-assisted UAV edge computing offloading decision and resource allocation optimization method as described above.

[0016] The beneficial effects of this invention are: 1. This invention achieves joint optimization of UAV deployment, unloading decisions and resource allocation, and adopts a dynamic resource allocation model. Resources are released and reallocated immediately after the task is completed, thereby improving resource utilization.

[0017] 2. This invention decomposes the complex mixed integer programming problem and creatively combines genetic methods and particle swarm optimization methods for alternating solutions, effectively solving the problem of coupling between discrete and continuous variables.

[0018] 3. This invention effectively avoids system bottlenecks through intelligent unloading decisions and refined resource scheduling, significantly reduces the average task completion time and user costs, and improves the overall system performance.

[0019] 4. This invention achieves optimized deployment of UAV physical locations through a two-stage process, laying a solid physical foundation for subsequent resource allocation and task unloading. This claim details how to determine the optimal hovering position of the UAV. First, it uses the computationally efficient K-Means algorithm to cluster users, quickly determining the initial deployment area of ​​the UAV. Then, through a refined local grid search, it finds the "golden spot" near the initial location that covers the most users. This ensures that the UAV is deployed in the location with the highest coverage from the outset, maximizing the system's service range and guaranteeing the utilization rate of system resources from a physical perspective.

[0020] 5. This invention, by establishing a mathematical model for caching decisions, achieves the strategic pre-configuration of service programs required for computational tasks, creating favorable conditions for low latency and high efficiency in subsequent task unloading. This claim defines the initial caching strategy for UAVs, introducing binary variables to describe "whether to cache a service" and clarifying the physical capacity limitations that caching must adhere to. This transforms caching decisions from a vague concept into a mathematical problem that can be precisely quantified and optimized. By pre-caching popular services, UAVs can quickly provide the service programs required for tasks, reducing the need to download services from base stations, which significantly reduces service acquisition latency.

[0021] 6. This invention constructs a precise, complete, and practically applicable multi-objective joint optimization problem, serving as a core bridge connecting "real-world problems" and "algorithm-based solutions." This claim formalizes and mathematically represents the optimization objective of the entire system. It defines the total system cost as the weighted sum of "task completion time" and "user cost," and details all relevant decision variables (caching, unloading, computing resources) and system constraints (caching capacity, computing power, etc.). Through weighting factors... The model also offers the flexibility to balance “efficiency” and “cost”, making it adaptable to different application scenarios.

[0022] 7. This invention decouples the complex discrete decision-making components of the original problem through problem decomposition, reducing the solution difficulty and achieving precise algorithm matching. This claim extracts the unloading decision in the joint optimization problem as an independent subproblem. In this subproblem, computational resource allocation is treated as a known parameter. This decoupling avoids blind searching in a large, mixed search space, greatly improving solution efficiency and solution quality.

[0023] 8. This invention decouples the continuous variable portion of the original problem, treating computational resource allocation as a separate subproblem. In this subproblem, the unloading decision is considered known. By decomposing the complex mixed-variable problem into purely discrete and purely continuous problems, this invention enables more efficient solutions to each subproblem.

[0024] 9. This invention proposes a collaborative solution framework that cleverly integrates the advantages of two different optimization algorithms, ensuring that the decomposed subproblems can work collaboratively to approach the global optimum. This claim defines how to solve the two subproblems iteratively and alternately. Specifically: first, fix the resource allocation and use a genetic algorithm to optimize the unloading decision; then fix the new unloading decision and use a particle swarm optimization algorithm to optimize the resource allocation, repeating this process until the total system cost no longer decreases significantly. It allows discrete and continuous decisions to "negotiate" and adapt to each other, gradually correcting each other's solutions. Through iteration, the solutions to the two subproblems continuously guide each other, gradually moving from a suboptimal solution to a high-quality joint optimum, effectively solving the difficulties caused by variable coupling when directly solving joint problems.

[0025] 10. This invention designs a unified and realistic fitness function, providing accurate and reliable guidance for two optimization algorithms. This evaluation is performed through an "event-driven simulator," which can dynamically calculate the total cost of the system under a given solution. The fitness function is the driving core of the heuristic algorithms (GA and PSO). By using an integrated simulator, dynamic processes such as task queuing and resource contention can be simulated more realistically, resulting in a more accurate evaluation of the solution and making the final optimized solution more realistically feasible and effective. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the multi-UAV edge computing system structure of the present invention; Figure 2 This is a comparative diagram showing the drone location deployment optimization before and after the present invention; Figure 3 This is a Gantt chart of dynamic computing resource allocation in time slot 1 according to the present invention; Figure 4 This is a Gantt chart of dynamic computing resource allocation in time slot 30 according to the present invention; Figure 5This is a comparison chart of the average system cost of the method of this invention and various benchmark methods under different numbers of users; Figure 6 This is a comparison chart of the average task completion time of the method of this invention and various benchmark methods under different numbers of users; Figure 7 This is a comparison chart of the average user cost of the method of this invention and various benchmark methods under different numbers of users. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1 A multi-drone edge computing system includes: one base station, multiple drones, and multiple ground users; the cache-assisted drone edge computing offloading decision and resource allocation optimization method includes the following steps: S1. Deploy multiple drones as edge servers, use a clustering method based on user location to determine the initial deployment location of each drone, and execute a local search strategy around the initial deployment location to maximize user coverage, and determine the final hovering location and initial cache of each drone. like Figure 1 As shown, the multi-UAV edge computing system in this embodiment consists of a base station, Unmanned Aerial Vehicles (UAVs) and It consists of ground users, where the ground user set drone collection Divide the total system time into Each time slot, in each time slot At the beginning, each user will be randomly generated. One task, of which The set of all tasks generated in each time slot is represented as ,in It is the first The total number of tasks generated per time slot, therefore Any task Both can be formed by triplets It means that, among them, Indicates task Input data size, Represents computational task Number of CPU cycles required Indicates the execution of a task The service types that the system depends on. The set of service types in the system is denoted as... ,therefore .

[0030] Each drone Equipped with a capacity of The drone can pre-store a portion of popular service programs using a binary variable. To represent drones Is the service cached? ,in Indicates service Cache on drones Up, or vice versa The caching decision must satisfy the drone's capacity constraint, which can be expressed as: (1) in, It is a service The size of the cache is determined by the different hotspot tasks in each drone's coverage area. During initialization, the drone caches services based on their popularity from high to low within capacity limits. If a task selects to download services from a base station during task execution, a cache replacement mechanism is triggered. According to the LFU (Least Frequently Used) policy, the least frequently accessed services in the cache are prioritized for removal, and the drone's cache content is dynamically updated.

[0031] The communication coverage area of ​​the drone is modeled as a circle with its horizontal position as the center and a radius of... A circular area, the radius of which is determined by the drone's flight altitude. and the half-power beamwidth angle of the airborne antenna A joint decision, specifically expressed as: (2) First, all users are divided into groups using the K-Means algorithm. non-overlapping clusters This minimizes the sum of squared Euclidean distances between users within a cluster, for each cluster. Calculate its geometric centroid and use it as the drone initial position To use drones The user coverage area is divided into a grid. All candidate locations in the grid are traversed, the number of users covered at each location is calculated, and the location with the highest coverage is selected as the final hovering location for the drone. Tasks created by users not covered by drones will be discarded.

[0032] This system employs three communication links: D2U (user-to-drone) communication, U2U (drone-to-drone) communication, and B2U (base station-to-drone) communication. Due to the presence of numerous buildings obstructing the urban environment, probabilistic line-of-sight (LoS) and non-line-of-sight (NLoS) channel models are used for the D2U and B2U links to more realistically reflect signal propagation characteristics. For the U2U communication link, a free-space path loss model is adopted. Furthermore, to improve spectral efficiency, both the D2U and U2U links utilize a hybrid NOMA mechanism.

[0033] For the link between the drone and the ground node, signal propagation is severely affected by building obstruction, and any transmitting node... and receiving node Its large-scale path loss Probability of line-of-sight transmission Decision, in the urban environment Depends on the elevation angle of the receiving node : (3) (4) in, It is a constant that depends on the type of environment. For the drone's altitude, The distance between nodes is the three-dimensional Euclidean distance. For the drone's altitude, for The angle between them This represents the line-of-sight transmission probability. Therefore, the corresponding non-line-of-sight probability is... It can be represented as (5) in, For line-of-sight transmission probability, This represents the probability of non-line-of-sight transmission. Average path loss It can be represented as: (6) (7) in, For free space path loss, For carrier frequency, and These represent the average additional losses under the LoS and NLoS states, respectively. Channel gain between It can be represented as: (8) Received power With transmission power The relationship can be represented as (9) Because drones fly at relatively high altitudes, line-of-sight (Line-of-Sight) links between drones are typically the dominant propagation path. Therefore, a free-space path loss model is used for U2U links. With drones Channel gain Defined as: (10) in, It is the channel gain at the reference distance. It is a drone With drones distance, This is the path loss index.

[0034] (1) D2U link When multiple users Simultaneously, the task is uploaded to the local drone. At that time, a packet-based hybrid NOMA access mechanism is adopted. Assume a drone... The user set of the service is To balance system complexity and receiver decoding performance, a head-to-tail pairing strategy is adopted for grouping. Specifically, the UAV first sorts all users from highest to lowest channel gain, pairing the "strong users" with the best channel conditions with the "weak users" with the worst channel conditions to form a NOMA group, and so on. Frequency division multiple access orthogonal transmission is used between groups. If the total bandwidth is... Then the bandwidth of each NOMA group is ,in This represents the total number of packets. NoMA transmission is used within each packet, and two users share the allocated sub-bandwidth. Assume that two users within a certain NoMA packet are both strong users. and weak users When the UAV decodes this group of signals, it uses SIC technology to first decode the strong user signals. At this point, the signals of the weak users are considered interference. Then, the strong user signals are subtracted from the received signals, and the weak user signals are decoded. At this point, the interference has been eliminated. Therefore, the transmission rate of users within the group can be expressed as: (11) (12) in, To increase the transmission rate for high-end users, For the transmission rate of weak users, It is the user's transmit power. and These are strong users and weak users with drones. Channel gain between It is the power of Gaussian white noise. For D2U channel bandwidth, This represents the total number of groups. Assume the task is uploaded to the drone. The transmission time is This time can be represented as follows: (13) in, The transmission time for uploading the task to the drone. It is a task Input data size, For users The transmission rate.

[0035] (2) U2U link Drones can transmit mission data and service programs to each other. When transmitting content to multiple neighboring drones, a packet-based downlink NOMA broadcast strategy is employed. To maximize the channel difference gain among users within the group, a head-to-tail pairing strategy is also used, with the source drone... Based on their distance from each target drone, the closest "near-end drone" and the farthest "far-end drone" are paired into a NOMA group, denoted as follows: and The transmitting end uses superposition coding to send signals. The near-end drone first decodes and cancels the far-end signal, while the far-end drone directly decodes its own signal. The transmission rate can be expressed as: (14) (15) in, The total number of groups, For the bandwidth allocated to each NOMA group, and These are the transmission powers of the near-end drone and the far-end drone, respectively. and These are near-end drones and long-end drones, and drones. Channel gain between It is the power of Gaussian white noise. For the U2U channel bandwidth, since the UAV flies at high altitudes and line-of-sight (LoS) links dominate, the channel gain is mainly determined by free-space path loss. Assume the UAV... Offload the task to the drone The transmission time is drones From drones The download time of the download service program is , respectively represented as: (16) (17) in, It is a task Required services Data size, For drones transmission rate For drones The transmission rate.

[0036] (3) B2U link When a drone downloads services through a base station, the link also follows a probabilistic line-of-sight model. Since the base station serves multiple drones simultaneously, bandwidth is allocated using orthogonal frequency division multiple access (OFDMA). Assuming the base station serves multiple drones simultaneously... The request is then assigned to the drone. The bandwidth is Download speed can be expressed as (18) in, It is the base station's transmission power. Base stations and drones Channel gain between.

[0037] Serve Download from base station to drone The time is: (19) When downloading services from a base station, users need to pay a fee to the base station, which is the user fee, denoted as . , can be represented as (20) in, It is the unit price for base station services per unit of data volume.

[0038] All calculations are performed on the drone; different drones have different computing capabilities, denoted as... Assuming the task... Choose drone Executed above, obtained The computing resources, then the task The computation delay can be expressed as: (twenty one) in, It is a task Calculate the required number of CPU cycles. It is worth noting that... It is not a static value; it refers to the total computing power of each drone. It is considered an infinitely divisible fluid resource and is allocated in real time proportionally based on the weights of all tasks concurrently executed on the drone. When the drone... Any task on When it completes its calculation and leaves the task queue, the computing resources it has consumed... The resources will be released immediately. The released resources will be instantly and automatically redistributed to the drones according to their weight ratios. All remaining tasks. Similarly, when a new task completes the communication phase, becomes computationally ready, and is added to the queue, the drone... The total computing resources will be reallocated proportionally across all tasks, including this new task.

[0039] S2, based on the final hovering position and initial cache of the UAV, constructs a joint optimization problem with the goal of minimizing the weighted sum of the average task completion time and average user cost (the average total cost of the system) of the multi-UAV edge computing system, and decomposes the joint optimization problem into a discrete offloading decision subproblem and a continuous computing resource allocation subproblem; use This represents the four unloading decisions for a task: Indicates task Local drones implement, Indicates task Unload from local drone to adjacent drone implement, Indicates task Local drones Execution, but downloading services via neighboring drones. Indicates task It is executed by a local drone, but the service is downloaded via a base station.

[0040] So, the task The processing time can be expressed as: (twenty two) in, For system task completion time, For user fees.

[0041] By optimizing UAV positioning, task offloading decisions, and resource allocation, the average total cost of the task is minimized, where the total system cost is... The weighted sum of task completion time and user fee is expressed as: (twenty three) in, Weighting factors for user fees.

[0042] Assume the total set of tasks completed within the entire time period is , The optimization problem is expressed as: (twenty four) in, These represent the drone caching decision, user offloading decision, and computing resource allocation strategy, respectively. express Drones The set of tasks to be executed on the drone; constraint C1 defines the drone's caching decision; constraint C2 indicates that the size of the drone's cache service cannot exceed the cache capacity limit; constraint C3 defines the user's offloading decision; constraint C4 indicates that the computing resources allocated to each task cannot exceed the total computing resource limit; constraint C5 indicates that if a task chooses to be offloaded to the local drone for execution, it must be cached on the local drone; constraint C6 indicates that if a task chooses to be offloaded to an adjacent drone for execution, the adjacent drone must cache the service; constraint C7 indicates that if a task chooses to download the service from an adjacent drone, the adjacent drone must cache the service.

[0043] The joint optimization problem P0 is decomposed into an unloading decision subproblem and a computational resource allocation subproblem. The goal of the unloading decision subproblem is to select an optimal unloading scheme for all newly arriving tasks, minimizing the average total system cost, while keeping the computational resource allocation weights temporarily fixed. Assume... To unload the decision set, the subproblem can be represented as: (25) All new task unloading decisions are encoded into an integer vector as a chromosome, with the value of each gene corresponding to a feasible unloading strategy index for a task.

[0044] The goal of the computational resource allocation subproblem is to find an optimal set of continuous computational resource allocation weights for all tasks that need to be processed, given that the unloading decision scheme is temporarily fixed, thus obtaining the computational resource allocation set for each task. To minimize the total system cost, the subproblem can be represented as follows: (26) in, Represents the set of computing resource allocations. For the total number of tasks, For a set of tasks, The total system cost, express Drones The set of tasks executed on top For drones computing power Indicates time slot drones For the task Allocated computing resources.

[0045] Allocate computational resources for all tasks to a weighted array and map it to a continuous real vector as the position of the particle, with each dimension having a value in the range (0,1).

[0046] S3 employs an alternating optimization strategy based on genetic methods and particle swarm optimization to solve the unloading decision subproblem and the computational resource allocation subproblem, and obtains the joint optimal solution through iteration; The process of solving the unloading decision subproblem and the computational resource allocation subproblem using an alternating optimization strategy includes: S31, Under a fixed weight of computing resource allocation, the genetic method (GA) is used to solve the unloading decision subproblem to obtain the optimal unloading decision combination; S32, fix the obtained unloading decision, use the particle swarm optimization method (PSO) to solve the computational resource allocation subproblem, and obtain the optimal computational resource allocation weight; S33, iteratively execute until the total system cost converges, and obtain the globally optimal joint scheme of unloading decision and computing resource allocation.

[0047] In this embodiment, both the genetic method and the particle swarm optimization method employ an integrated event-driven simulator as the fitness function. The simulator evaluates and outputs the corresponding total system cost based on the input unloading decisions and resource allocation schemes. The fitness function comprises two modules: communication and computation, as detailed below: (1) Communication delay model For a given unloading decision, the function first identifies the concurrent data streams of all new tasks. Based on a hybrid NOMA model and a head-to-tail pairing strategy, it computes the data for each new task. Total communication latency: (27) in, These are the transmission delays of D2U, U2U, and B2U links obtained based on the hybrid NOMA rate model.

[0048] (2) Calculation of delay model The function uses a generalized processor-shared model to simulate the computation process of all incomplete tasks. For drones... Task queue on Assuming drones The computing resources are ,Task Resource weight So, the task The allocated computing resources can be represented as: (28) The above formula shows that as long as a task is being executed, the drone's computing resources are fully utilized. When a task completes and leaves the queue, the resources it occupied are immediately redistributed to the remaining tasks according to the new weight ratio. Assume the task... The remaining computational workload is So, the task The computation time required is: (29) (3) Fitness function Define the fitness function as the total cost required to process a task, for the old task set. and new task collection The fitness function is: (30) in, As a weighting factor for user fees, For user fees, This is the final fitness value of the joint scheme.

[0049] The solution is obtained using an alternating iterative optimization framework based on genetic methods (GA) and particle swarm optimization (PSO); specifically, in each iteration: By fixing the weights of computing resource allocation, the unloading decision subproblem is mapped to an optimization problem using genetic methods. Through operations such as selection, crossover, and mutation, the optimal combination of task unloading schemes is searched. Given the optimal unloading scheme found in the previous step, the resource allocation subproblem is transformed into an optimization problem using the particle swarm optimization method. By updating the velocity and position of the particles, the optimal resource allocation weight vector is found. This process iterates repeatedly until the total system cost converges, ultimately yielding a high-quality joint optimization solution.

[0050] After the alternating optimization is complete, for tasks that decide to download services from the base station, the system will trigger the LFU (Least Recently Used) cache replacement mechanism. If the target drone's cache is full, the least frequently used service will be evicted to make room for the new service.

[0051] Based on a similar inventive concept, embodiments of the present invention also provide a computer storage medium storing a readable program that, when run by a processor, can execute the aforementioned cache-assisted UAV edge computing offloading decision and resource allocation optimization method.

[0052] Based on a similar inventive concept, this invention provides an electronic device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the cache-assisted UAV edge computing offloading decision and resource allocation optimization method described above.

[0053] Based on a similar inventive concept, embodiments of the present invention also provide a computer program product, including computer instructions, which instruct a computing device to perform operations corresponding to the cache-assisted UAV edge computing offloading decision and resource allocation optimization method described above.

[0054] Example 2 In this embodiment, a cache-assisted UAV edge computing offloading decision and resource allocation optimization device is proposed, specifically including: Location confirmation module: Deploy multiple drones as edge servers, use a clustering method based on user location to determine the initial deployment location of each drone, and execute a local search strategy around the initial deployment location to maximize user coverage, and determine the final hovering location and initial cache of each drone. Problem Construction and Decomposition Module: Based on the final hovering position and initial cache of the UAV, a joint optimization problem is constructed with the goal of minimizing the average total cost of the multi-UAV edge computing system, and the joint optimization problem is decomposed into a discrete offloading decision subproblem and a continuous computing resource allocation subproblem; Problem-solving module: It uses an alternating optimization strategy based on genetic methods and particle swarm optimization to solve the unloading decision subproblem and the computational resource allocation subproblem, and obtains the joint optimal solution through iteration.

[0055] Example 3 In this embodiment, the performance of the method of the present invention is compared and verified with multiple benchmark methods under different numbers of users. The verification process includes: simulation experiments implemented using the MATLAB platform. The scenario is set in a... Six drones were deployed within the square area. Provide services to ground users by default. The altitude of the drone Each user per time slot Randomly generated There are 1 computational task, and the amount of data for each task is... The values ​​are randomly generated, and the number of CPU cycles required for each task's computation is within a certain range. The tasks are randomly generated, and their popularity follows a Zipf distribution. Other key simulation parameters are shown in Table 1.

[0056] Table 1 Simulation Parameters To verify the performance of the proposed GA-PSO algorithm, it is compared with the following three baseline algorithms: (1) Dynamic Equal Algorithm (DE) This algorithm decouples unloading decisions from resource allocation. First, it uses only a genetic algorithm (GA) to search for the optimal unloading decision for all newly arriving tasks. In the subsequent task execution phase, the computing resources of each drone are evenly distributed among all tasks concurrently executing on it. When a task is completed, the resources are dynamically redistributed to the remaining tasks.

[0057] (2) Random Offload (RO) Algorithm This algorithm can verify the importance of resource allocation optimization. For each newly arrived task, its offloading decision is randomly selected from all its feasible strategies. After the offloading decisions for all tasks have been randomly determined, the algorithm calls the particle swarm optimization algorithm to search for the optimal computational resource allocation weights.

[0058] (3) Random NOMA pairing algorithm (R-NOMA) This algorithm is a variant of the proposed algorithm, used to verify the effectiveness of the head-to-tail pairing strategy in NOMA. It employs the exact same GA-PSO alternating optimization algorithm as the algorithm in this paper to jointly optimize offloading decisions and resource allocation. However, during NOMA transmission, it groups users or UAVs randomly, rather than pairing them head-to-tail based on channel quality.

[0059] Figure 2The diagram presents the deployment and user coverage status of drones before and after location optimization. In the figure, the pentagram represents the base station location, situated at the center of the area. The black cross represents the initial hovering position of the drone, determined solely by the K-Means clustering algorithm; this position is the geometric centroid of all users within its cluster. The colored triangle represents the final hovering position of the drone after the grid-search-based maximum coverage optimization algorithm. As can be seen from the diagram, the optimized drone position shifts compared to its initial position. This is because the optimization algorithm no longer aims to minimize the average distance to all users, but rather to find the optimal hovering position that includes the most users within its circular coverage area. For example, for the drone in the green area, its initial centroid is close to the area center, but the optimized position moves towards a more densely populated area, thus bringing users at the edge of its coverage area into service. The dark dots in the diagram represent users ultimately covered by the drone, while the semi-transparent dots represent users whose tasks were dropped due to being outside the coverage area. This optimization process effectively improves user coverage, thereby reducing the system's task drop rate.

[0060] The verification results are as follows Figures 3 to 7 As shown; Figure 3 and Figure 4 Two time slots in the simulation process were selected to create a Gantt chart of UAV computational resource allocation. In the chart, the horizontal axis represents the simulation time, the vertical axis represents the UAV's computational resources, and the red dashed line represents the upper limit of the UAV's computational capacity. Each colored stacked block represents the amount of computational resources obtained by an independent task within a specific time period. Figure 3 The resource allocation from 0 to 1 second is shown in the figure. As can be seen, the system load is low during this phase, and the number of concurrent tasks executed on each UAV is small. Because tasks need to go through a communication phase first, most tasks only gradually become computationally ready and begin execution later. When a task runs alone, it can obtain all the UAV's computing resources to accelerate its completion; for example, blue task 11 in UAV1 occupies all computing resources. When a new task joins, the algorithm immediately reallocates resources. For example, in UAV2, tasks 50 and 86: when task 50 is not ready, task 86 temporarily monopolizes computing resources; when task 50 becomes ready, computing resources are reallocated according to optimized weights; and when task 50 completes, computing resources are released back to task 86 for processing, demonstrating its rapid response and dynamic scheduling capabilities. Figure 4The resource allocation during 29-30 seconds is shown in the graph. As can be seen, the system load is high at this time, with more than 15 tasks running simultaneously on multiple drones. These include unfinished tasks carried over from the previous time slot and short tasks arriving in the current time slot. The graph also shows that when a task completes, its resources are immediately taken over by other running tasks, keeping the drone's computing power consistently saturated, demonstrating the work-conserving characteristic of the algorithm. Similarly, when a new task becomes ready and joins the computation queue, the resources of existing tasks are dynamically compressed to accommodate the new task.

[0061] Figure 5 The average weighted cost of the system as a function of the number of users was plotted. As can be seen from the figure, the average system cost of the proposed GA-PSO algorithm is significantly lower than all benchmark algorithms across all user scales, demonstrating the effectiveness of joint optimization of offloading decisions and resource allocation using GA-PSO. Specifically, GA-PSO shows the most significant performance advantage compared to the RO algorithm, highlighting the crucial role of offloading decisions in joint optimization. Random decisions can easily lead to tasks being assigned to drones with congested computing or communication links, or to choosing inefficient service acquisition paths, resulting in severe queuing and resource conflicts, leading to high overall system costs. Compared to the DE algorithm, GA-PSO also demonstrates superior performance. Both algorithms employ GA to optimize offloading decisions; the main difference lies in the allocation of computational resources. GA-PSO uses PSO to find differentiated resource allocation weights for different tasks, while DE uses a simple equal-distribution strategy. The results show that PSO can perform more refined resource scheduling based on task heterogeneity, thereby achieving lower system costs. Figure 6 The relationship between average task completion time and the number of users was plotted. The GA-PSO algorithm effectively avoids system bottlenecks through intelligent offloading decisions and resource allocation, enabling tasks to be assigned to the most suitable UAVs for rapid processing, thus achieving the lowest average completion time. Compared with the R-NOMA algorithm, GA-PSO also has significant advantages. This proves the effectiveness of the head-to-tail pairing strategy adopted in this paper's NOMA communication model. Head-to-tail pairing maximizes the channel gain difference among users within the NOMA group, making serial interference cancellation (SIC) at the receiver more effective, thereby achieving higher spectral efficiency and lower communication latency. Figure 7The graph shows the average user cost as a function of the number of users. As can be seen, the GA-PSO algorithm also performs best in controlling user costs. GA-PSO, through its global search capability, maximizes the distributed caching of the drone swarm, prioritizing free services from local or neighboring drones and only downloading from base stations for a fee when necessary. In contrast, the RO algorithm, due to the randomness of its decision-making, results in a large number of unnecessary base station downloads, thus producing the highest average user cost. While the DE and R-NOMA algorithms are superior to RO, their suboptimal resource allocation or communication efficiency makes their ability to balance time and cost weaker than GA-PSO, ultimately resulting in higher costs.

[0062] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0063] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A cache-assisted method for unloading decision-making and resource allocation optimization in drone edge computing, wherein the multi-drone edge computing system includes: A base station, multiple drones, and multiple ground users, characterized in that the method includes the following steps: Multiple drones are deployed as edge servers. A clustering method based on user location is used to determine the initial deployment location of each drone. A local search strategy is then executed around the initial deployment location to maximize user coverage and determine the final hovering location and initial cache of each drone. Based on the final hovering position and initial cache of the UAV, a joint optimization problem is constructed with the goal of minimizing the average total cost of the multi-UAV edge computing system. The joint optimization problem is decomposed into a discrete offloading decision subproblem and a continuous computing resource allocation subproblem. An alternating optimization strategy based on genetic methods and particle swarm optimization is used to solve the unloading decision subproblem and the computational resource allocation subproblem, and the joint optimal solution is obtained through iteration.

2. The cache-assisted UAV edge computing offloading decision and resource allocation optimization method according to claim 1, characterized in that, The process of determining the final hovering position of each drone includes: All users are divided into groups using the K-Means algorithm. non-overlapping clusters This minimizes the sum of squared Euclidean distances between users within a cluster, for each cluster. Calculate its geometric centroid and use it as the drone initial position ; drones The user coverage area is divided into a grid. All candidate locations in the grid are traversed, the number of users covered at each location is calculated, and the location with the highest coverage is selected as the final hovering location for the drone. Tasks generated by users not covered by drones will be discarded.

3. The cache-assisted UAV edge computing offloading decision and resource allocation optimization method according to claim 2, characterized in that, The process of determining the initial cache is as follows: Each drone Equipped with a capacity of The drone pre-stores a portion of popular service programs in its cache space using a binary variable. Indicates drone Is the service cached? ,in Indicates service Cache on drones Up, or vice versa Cache decisions must satisfy the drone's capacity constraints, expressed as: in, It is a service The size of the drone set is K, where K represents the drone set.

4. The cache-assisted UAV edge computing offloading decision and resource allocation optimization method according to claim 1, characterized in that, The joint optimization problem is: in, These represent the drone caching decision, user offloading decision, and computing resource allocation sets, respectively. express The set of tasks at any given moment; For the total number of tasks, For a set of tasks, For caching decisions, It is a service The size of K represents the drone set. For unloading decision set, For drones computing power Indicates time slot drones For the task Allocated computing resources; The total system cost, As a weighting factor for user fees, For system task completion time, For user fees; For the latency of user uploads to the drone, For the computational latency of the drone, For the mission from drone Transmit to drone The time delay, For the service program from Transmit to The time delay, This refers to the latency of downloading service programs from the base station.

5. The cache-assisted UAV edge computing offloading decision and resource allocation optimization method according to claim 4, characterized in that, The unloading decision subproblem is: in, These represent the user uninstallation decision set, For the total number of tasks, For a set of tasks, The total system cost, express The set of tasks at any given moment; For the unloading decision set.

6. The cache-assisted UAV edge computing offloading decision and resource allocation optimization method according to claim 4, characterized in that, The computational resource allocation subproblem is: in, Represents the set of computing resource allocations. For the total number of tasks, For a set of tasks, The total system cost, express Drones The set of tasks executed on top For drones computing power Indicates time slot drones For the task Allocated computing resources.

7. The cache-assisted UAV edge computing offloading decision and resource allocation optimization method according to claim 1, characterized in that, The process of solving the unloading decision subproblem and the computational resource allocation subproblem using an alternating optimization strategy includes: S31, Under a fixed weight of computing resource allocation, the genetic method is used to solve the unloading decision subproblem to obtain the optimal unloading decision combination; S32, fix the obtained unloading decision, use the particle swarm optimization method to solve the computational resource allocation subproblem, and obtain the optimal computational resource allocation weight; S33, iteratively execute until the total system cost converges, and obtain the globally optimal joint scheme of unloading decision and computing resource allocation.

8. The cache-assisted UAV edge computing offloading decision and resource allocation optimization method according to claim 7, characterized in that, Both the genetic method and the particle swarm optimization method use an integrated event-driven simulator as the fitness function; the fitness function expression is: in, For the fitness function, Represents the set of old tasks. Represents a new set of tasks. For the task The computational latency required, For each new task Total communication latency, As a weighting factor for user fees, For user fees, This is the final fitness value of the joint scheme; For the task The remaining computational load, Indicates drone The task queue on the screen, Indicates drone Computing resources For the task Resource weight, Indicates task The allocated computing resources; These are the transmission delays of D2U, U2U, and B2U links obtained based on the hybrid NOMA rate model.

9. A cache-assisted UAV edge computing offloading decision and resource allocation optimization device, executing the method according to any one of claims 1-8, characterized in that, include: Location confirmation module: Deploy multiple drones as edge servers, use a clustering method based on user location to determine the initial deployment location of each drone, and execute a local search strategy around the initial deployment location to maximize user coverage, and determine the final hovering location and initial cache of each drone. Problem Construction and Decomposition Module: Based on the final hovering position and initial cache of the UAV, a joint optimization problem is constructed with the goal of minimizing the average total cost of the multi-UAV edge computing system, and the joint optimization problem is decomposed into a discrete offloading decision subproblem and a continuous computing resource allocation subproblem; Problem-solving module: It uses an alternating optimization strategy based on genetic methods and particle swarm optimization to solve the unloading decision subproblem and the computational resource allocation subproblem, and obtains the joint optimal solution through iteration.

10. A computer storage medium storing a readable program, characterized in that, When the program runs, it can instruct the computing device to execute the cache-assisted UAV edge computing offloading decision and resource allocation optimization method as described in any one of claims 1-8.

Images