An adaptive cooperative air-ground edge caching method in a dynamic vehicle networking scene

Abstract

The application discloses a kind of dynamic car networking scene under adaptive cooperative air-ground edge caching method, this method first is to establish a dynamic edge caching system, then consider the dynamic request of vehicle user in multilayer collaborative network, construct dynamic edge caching model, and calculate service delay and caching cost, and then the caching resource allocation between UAV and RSU is designed as a joint optimization problem, to minimize the total cost of edge task delivery, the two indexes of service delay and caching cost are balanced in air-ground cooperative edge caching architecture;Finally, the dynamic edge caching model is optimized and solved based on the ACCO algorithm of kernel search heuristic framework and greedy strategy, so as to realize the allocation of edge caching task on unmanned aerial vehicle, roadside unit or base station.

Description

Technical Field

[0001] This invention pertains to edge computing tasks, specifically relating to an adaptive cooperative air-to-ground edge caching method in dynamic vehicle-to-everything (V2X) scenarios. Background Technology

[0002] In recent years, the automotive industry has seen a continuous increase in the trend towards intelligence and networking, leading to greater attention being paid to in-vehicle entertainment systems, of which real-time video files are a crucial element. These files are a major component of the entertainment experience and therefore require seamless transmission with minimal latency. Furthermore, the massive data transmission volumes place enormous pressure on vehicle networks that rely entirely on cellular networks.

[0003] Mobile edge caching is a key technology in sixth-generation (6G) networks because it effectively addresses backhaul burdens and reduces round-trip latency. With decreasing storage costs, implementing caching at the wireless edge is becoming increasingly cost-effective. However, due to the unique characteristics of vehicular networks, such as fast-moving dynamic network topologies and high vehicle speeds, supporting Intelligent Transportation Systems (ITS) services with only one wireless communication technology is challenging. To address this issue, heterogeneous vehicular networks (H-VNets) are considered a future solution for vehicle users. In H-VNets, base stations (BSs) and roadside units (RSUs) can collaborate to cache content, significantly reducing vehicle communication costs. Furthermore, unmanned aerial vehicles (UAVs) have recently attracted considerable interest because they can serve as edge servers to provide services to users in certain scenarios due to their low cost, flexible deployment, and high mobility. For example, UAVs and ground BSs can cache video during off-peak hours and serve mobile users during peak hours. This flexible deployment of UAVs compensates for the limited coverage provided by ground infrastructure. Meanwhile, future ITS are expected to enable "cinemas on wheels" applications, allowing mobile users to access video and audio content during peak traffic periods. This highlights the importance of developing new applications supported by edge caching servers. Therefore, collaborative edge caching in H-VNets is crucial for future vehicular networks, reducing network congestion and improving the user's driving experience. Unfortunately, existing research has not explored how to optimize cache placement to facilitate continuous communication between vehicles and infrastructure in dynamic traffic scenarios. Furthermore, previous research has not integrated UAVs, RSUs, and BSs to collaboratively provide delivery services to vehicles within the vehicular network. Finally, the simultaneous optimization of service latency and caching costs remains a fundamentally unexplored aspect of mobile edge caching.

[0004] Meanwhile, mobile edge caching is a promising technology supporting Intelligent Transportation Systems (ITS) applications, which involve large amounts of image and video content. Specifically, key challenges arise in meeting the content requirements of mobile vehicles. In this work, the invention investigates collaborative content delivery in an integrated air-ground vehicle network, where unmanned aerial vehicles (UAVs) and roadside units (RSUs) cache popular content and further deliver it to nearby vehicles. Considering the high mobility of vehicles and UAVs, and the limited caching capacity of edge nodes, designing an effective content placement strategy to minimize service latency for vehicles and caching costs for edge nodes is crucial. Summary of the Invention

[0005] Purpose of the invention: This invention provides an adaptive cooperative air-ground edge caching method for dynamic vehicle-to-everything (V2X) scenarios, enabling dynamic allocation of edge caches in V2X scenarios.

[0006] Technical solution: An adaptive cooperative air-to-ground edge caching method for dynamic vehicle-to-everything (V2X) scenarios, comprising the following steps:

[0007] (1) Construct a dynamic edge caching system to dynamically network vehicles, unmanned aerial vehicles (UAVs), roadside units (RSUs), and base stations (BSs) that have edge computing task caching capabilities;

[0008] (2) Establish channel models between UAV, RSU, BS and vehicles, including U2V channel model, R2V channel model and B2V channel model, and calculate the corresponding transmission rates as follows:

[0009] The transmission rate between the unmanned aerial vehicle u and the vehicle v is calculated as follows:

[0010]

[0011] Among them B u,v and P u,v These represent the bandwidth and transmission power between UAV u and vehicle v, respectively. Indicates noise power;

[0012] The transmission rate from roadside unit r to vehicle v is expressed as:

[0013]

[0014] Among them B r,v and P r,v These represent the communication bandwidth and transmit power between the RSU and the vehicle, respectively. This indicates the receiver noise power, including potential interference and receiver thermal noise;

[0015] The achievable content transfer rate from base station BS b to vehicle v is expressed as follows:

[0016]

[0017] Among them B b,v and P b,v These represent the communication bandwidth and transmission power between the BS and the vehicle, respectively. Indicates the noise power of the receiver;

[0018] (3) Establish a dynamic edge caching model, and consider service latency and caching costs;

[0019] definition and Place vectors for UAV u and RSU r in each content and time slot, where u represents the unmanned aerial vehicle (UAV), f represents the cache task, t represents the time slot, and r represents the roadside unit.

[0020] Where α u,f,t ,β r,f,t ∈{0,1}; In addition, the total size of the cached content in each UAV and RSU should be within the available cache capacity Qu and Qr, respectively, where Qu represents the available cache capacity of the unmanned aerial vehicle and Qr represents the available cache capacity of the roadside unit;

[0021] (4) The allocation of cache resources between UAV and RSU is designed as a joint optimization problem to minimize the total cost of edge task delivery. In the air-ground cooperative edge caching architecture, the two indicators of service latency and cache cost are weighed.

[0022] The optimization problem is limited to the Multidimensional Multiple Knapsack Problem (MMKP), and its mathematical formulation is as follows:

[0023] P1:

[0024] st(a)α u,f,t ,β r,f,t ∈{0,1},

[0025] (b)

[0026] (c)

[0027] (d)

[0028] (e)

[0029] Where ∈{0,1} is a weighted parameter, constraint (a) is a binary variable constraint representing the indivisibility of content; constraint (b) stipulates that a vehicle can only be served by one edge node at a time; constraint (c) limits the maximum amount of content that can be cached in UAVs and RSUs; constraint (d) stipulates the maximum allowed connection time between UAVs, RSUs and vehicles;

[0030] (5) The ACCO algorithm based on the kernel search heuristic framework and greedy strategy is used to optimize the dynamic edge caching model.

[0031] Furthermore, in step (1), the UAV flies over the area covered by the vehicle-to-everything (V2X) system to provide flexible connectivity, and the RSU is deployed on the road to provide stable connectivity; and each UAV and RSU is connected to an edge cache server; the BS is a cloud server that contains all requested content; the UAV, RSU and BS communicate with the vehicle through a cellular vehicle-to-everything (V2X) interface.

[0032] In a dynamic edge caching system, the unmanned aerial vehicle U, roadside unit R, base station B, and vehicle V are represented as U = {1,2,...,U}, R = {1,2,...,R}, B = {1,2,...,B}, and V = {1,2,...,V}.

[0033] Due to the high dynamism, let Lu and Lr represent the finite contact time between UAV u∈U, RSU r∈R, and vehicle v∈V, respectively; BS b∈B provides a stable link for the vehicle, and let the set F = {1, 2, ..., F} represent the index set of available content, where the size of content f∈F is defined as z. f The content size vector formed by the sizes of all content is Z = [z1, z2, ..., z]. f ,...,z F Furthermore, the entire timeline is divided into T time periods of equal length, with each time period indexed by t.

[0034] Furthermore, step (2) for calculating the edge task transmission rate under the U2V channel model, R2V channel model, and B2V channel model includes the following considerations:

[0035] For the U2V channel model, considering both line-of-sight and non-line-of-sight aspects of the U2V channel, its expression is as follows:

[0036]

[0037]

[0038] Where f u,v and L represents the carrier frequency and the distance from UAV u to vehicle v. LoS and L NLoS The shading and diffraction losses are for LoS and NLoS, respectively.

[0039] For UAVs and vehicles, the probabilities of LoS and NLoS links depend on their altitude and environment, and can be expressed as:

[0040]

[0041] p NLoS =1-p LoS ,

[0042] Where a and b are constants related to the environment, and θ is the angle of elevation in radians;

[0043] Therefore, the average path loss associated with LoS and NLoS is expressed as follows:

[0044]

[0045] For the R2V channel model, assuming the time-varying attenuation channel follows a Rician distribution, the expression for calculating the channel gain is as follows:

[0046]

[0047] in A is a constant. d and d e f represents the antenna gain and path loss exponent, respectively. r,v For the carrier frequency, d r,v This represents the distance from RSU r to vehicle v;

[0048] For the B2V channel model, considering the small-scale attenuation effect in B2V communication, the channel gain is calculated as follows:

[0049]

[0050] Where g0 and d0 represent the path loss constant and reference distance, respectively. It is the distance from BS to the vehicle.

[0051] Furthermore, in step (3), for the dynamic edge caching model, according to the Zipf distribution, the popularity of edge caching task f is represented as follows:

[0052]

[0053] j represents the Zipf index, indicating the skewness of popularity, and ξ represents the ranking of content request frequency.

[0054] The service latency of the dynamic edge caching model described in step (3) is represented as follows:

[0055]

[0056] The cost of caching is described as:

[0057]

[0058] ρ u ρ r and ρ b N represents the unit cost of renting cache resources from UAV u, RSU r, and BS b, respectively. u N r and N b S represents the number of edge nodes per unit area, i.e., UAV, RSU, and BS. u S r and S b . These represent the number of edge cache tasks cached on UAV u, RSU r, and BS b, respectively.

[0059] Furthermore, step (5) for solving the dynamic edge caching model includes the following three stages:

[0060] Initialization phase: Relax the linear programming of the original problem, use the k-means clustering algorithm to classify the content, and obtain an upper bound;

[0061] Iterative solution process: The joint optimization problem of cache resource allocation between UAV and RSU is decomposed into a series of sub-problems, and then the CPLEX tool is used to iteratively solve the problem to obtain a feasible solution;

[0062] Dynamic adjustment phase: The dimensions and solution time of the subproblems are dynamically adjusted, specifically by using a greedy load balancing strategy to further optimize the solution results.

[0063] Furthermore, the initialization phase includes optimizing the constraints (a) of the objective function to obtain the linear programming relaxation of the original problem P1, which is then transformed into solving problem P2. The mathematical representation of problem P2 is as follows:

[0064] P2:

[0065] st(a) 0 ≤ α u,f,t ,β r,f,t ≤1,

[0066] (b)~(e)

[0067] For problem P2, the K-means clustering algorithm minimizes the sum of squares within each cluster, i.e., the square of the Euclidean distance, and the objective expression is as follows:

[0068]

[0069] Where μ k Is the user in cluster s k The average position in;

[0070] The iterative solution process involves transforming problem P2 into problem P3, mathematically represented as follows:

[0071] P3:

[0072] st(a)

[0073] (b)

[0074] m represents the number of items contained in each subproblem, which will be dynamically adjusted in later stages; and It is an adjusted fixed value. and Represented as:

[0075]

[0076]

[0077] Therefore, Λ∪BK is constructed. i The set can be solved using the CPLEX tool by adding the following two constraints.

[0078]

[0079] z cut =z * +1

[0080] Where z cut It is the cutoff value of the objective function of the problem;

[0081] The dynamic adjustment phase includes designing a greedy load balancing strategy to select items to include in the kernel; assuming different vehicles might request the same content in the same region, the value ratio e involved in this strategy is adjusted accordingly. f,u and e f,r Defined as:

[0082]

[0083]

[0084] in W is a weighted parameter. f This indicates the request frequency of content f. If e f,u ,e f,r >1, the vehicle will request BS; if e f,u ,e f,r If e < 1, then these values ​​are compared; if e < 1, then e < 1. f,u <e f,r If the content is cached on the UAV, then the content is cached on the RSU.

[0085] Beneficial effects: The significant effects and substantial features of the method described in this invention include the following:

[0086] This invention first establishes a dynamic edge caching system and proposes an air-ground cooperative edge caching architecture, enabling vehicles to collaboratively request content from UAVs, RSUs, and BSs within a dynamic vehicle network. Next, the invention formulates the heterogeneous caching optimization problem as a generalized MMKP, minimizing service latency by achieving seamless content delivery and reducing caching costs by storing content on low-cost servers. Furthermore, the invention incorporates an Adaptive Collaborative Cache Optimization (ACCO) algorithm in solving the model, which utilizes a kernel search (KS) framework to address subproblems with dynamically adjustable dimensions and solution time. Additionally, a k-means clustering algorithm is used to classify content based on service latency, caching cost, and content size, and a greedy load balancing strategy is added to effectively match content with edge nodes. Finally, extensive simulations demonstrate the strong effectiveness of the proposed caching architecture and algorithm in minimizing service latency and caching costs compared to other classic algorithms. Attached Figure Description

[0087] Figure 1 This is a schematic diagram of the system constructed in this invention;

[0088] Figure 2 This is a flowchart illustrating the ACCO algorithm;

[0089] Figure 3 This is a graph analyzing service latency and caching costs for six algorithms under different request volumes;

[0090] Figure 4 This is a graph showing the impact of the Zipf parameter j of the popularity distribution on service latency and caching costs.

[0091] Figure 5 It represents the percentage of service latency and caching cost for different algorithms across 1000 requests;

[0092] Figure 6 This is a comparison chart of the runtime of different algorithms for a certain range of request counts. Detailed Implementation

[0093] To illustrate the technical solution disclosed in this invention in detail, a specific description is provided below.

[0094] Mobile edge caching is a promising technology supporting Intelligent Transportation Systems (ITS) applications involving large amounts of image and video content. Specifically, key challenges arise in meeting the content requirements of mobile vehicles. This invention first investigates collaborative content delivery in an integrated air-ground vehicle network, where unmanned aerial vehicles (UAVs) and roadside units (RSUs) cache popular content and further deliver it to nearby vehicles. Considering the high mobility of vehicles and UAVs, and the limited caching capacity of edge nodes, designing an effective content placement strategy to minimize service latency for vehicles and caching costs for edge nodes is crucial. To this end, the heterogeneous caching optimization problem is formulated as a generalized multidimensional multi-node packet problem (MMKP), which has been proven to be NP-hard. In particular, this invention also proposes an Adaptive Collaborative Caching Optimization (ACCO) algorithm, which employs a kernel search (KS) framework to find a near-optimal caching scheme, with dynamically adjustable subproblem dimensions and solution time. Furthermore, the k-means clustering algorithm is used to classify content based on service latency, caching cost, and content size, combined with a greedy load balancing strategy to achieve a cost-effective optimal match between content and edge nodes. Comprehensive simulations validate the effectiveness of the proposed collaborative edge caching architecture and ACCO algorithm in reducing both service latency and caching costs during content distribution.

[0095] Furthermore, the air-ground collaborative edge caching architecture constructed in this invention allows UAVs, RSUs, and BSs to cooperate in caching and delivering popular content to moving vehicles within a dynamic vehicle network. Considering cross-layer cooperation and vehicle mobility, the travel time of vehicles within the coverage areas of UAVs and RSUs was first analyzed, which determined the upper limit of the content delivery time for a single node. To minimize service latency and the caching cost of content delivery, a caching optimization problem was subsequently formulated for content placement design: the generalized multidimensional multiple knapsack problem (MMKP). In this problem, each edge node is considered a unique two-dimensional knapsack, with cached content as items.

[0096] 1. Construct a dynamic edge caching system to dynamically network vehicles, unmanned aerial vehicles (UAVs), roadside units (RSUs), and base stations (BSs) that have edge computing task caching capabilities;

[0097] like Figure 1 As shown, Figure 1The proposed air-to-ground collaborative edge caching scenario is illustrated. The scenario includes several UAVs, RSUs, and a BS (Browser Base Station). Vehicles traveling within this area may request popular content such as news and video services. UAVs fly over the area to provide flexible connectivity, while RSUs are deployed on the roads to provide stable connectivity. Each UAV and RSU connects to an edge caching server, which caches a limited amount of popular content for nearby vehicles. Additionally, the BS acts as a cloud server, containing all requested content and maintaining robust communication capabilities. The UAVs, RSUs, and BS communicate with the vehicles via a cellular vehicle-to-everything (C-V2X) interface.

[0098] Table 1

[0099] UAV Lower Cheap Remote, dynamic RSU Medium Right Medium distance BS High Expensive Near, reliable

[0100] Table 1 briefly summarizes the characteristics of different edge nodes. Due to their high deployment and maintenance costs, RSUs and BSs are considered to have higher content caching costs than UAVs. However, UAVs have the lowest transmission rate due to inter-link interference. In edge caching applications, vehicle users may frequently switch between different caching servers to ensure seamless communication. Therefore, this invention employs cooperative content delivery in the vehicle network to meet the requirements of latency and caching cost. Different spectrum frequencies are used for UAV-to-vehicle (U2V), RSU-to-vehicle (R2V), and BS-to-vehicle (B2V) communication to avoid co-channel interference.

[0101] Furthermore, consider U UAVs, R RSUs, B BSs, and V vehicles, which are contained in sets U = {1,2,...,U}, R = {1,2,...,R}, B = {1,2,...,B}, and V = {1,2,...,V}, respectively. Due to the high dynamism, let Lu and Lr represent the finite contact time between UAVs u∈U, RSUs r∈R, and vehicles v∈V, respectively. BS b∈B provides a stable link for the vehicles. F = {1,2,...,F} represents the index set of available content. The size of the content f∈F is defined as z. f The content size vector formed by the sizes of all content is Z = [z1, z2, ..., z]. f ,...,z F Furthermore, the entire timeline is divided into T time periods of equal length, with each time period indexed by t.

[0102] (2) Channel models between vehicles and various types of edge nodes, assuming that the environmental state remains unchanged within a time slot. Establish channel models between UAVs, RSUs, BSs and vehicles, including U2V channel models, R2V channel models and B2V channel models.

[0103] (21) U2V channel model

[0104] Signals from all wireless channels from the UAV to the vehicle are likely to be attenuated due to strong reflection and diffraction. Therefore, a path loss model is used, which takes into account both line-of-sight (LoS) and non-line-of-sight (NLS) aspects of the U2V channel. Its expression is as follows:

[0105]

[0106]

[0107] Where f u,v and This represents the carrier frequency and the distance from UAV u to vehicle v. L LoS and L NLoS The shading and diffraction losses are for LoS and NLoS, respectively.

[0108] For UAVs and vehicles, the probabilities of LoS and NLoS links depend on their altitude and environment, and can be expressed as:

[0109]

[0110] p NLoS =1-p LoS ,

[0111] Where a and b are constants related to the environment, and θ is the angle of elevation in radians.

[0112] Accordingly, the average path loss associated with LoS and NLoS can be derived as follows:

[0113]

[0114] Then the transmission rate between UAV u and vehicle v can be obtained, and its formula is:

[0115]

[0116] Among them B u,v and P u,v These represent the bandwidth and transmission power between the UAV (u) and the vehicle (v), respectively. Indicates noise power.

[0117] (22) R2V channel model

[0118] In densely built-up urban areas, radio signals are severely dispersed. Assume the time-varying attenuation channel follows a Rician distribution. The channel gain is given by the following formula:

[0119]

[0120] Where θ is a constant. A d and d e These represent the antenna gain and path loss exponents, respectively. r,v For carrier frequency, dr ,v This represents the distance from RSU r to vehicle v.

[0121] The achievable transfer rate from RSU r to vehicle v can be expressed as:

[0122]

[0123] Among them B r,v and P r,v These represent the communication bandwidth and transmission power between the RSU and the vehicle, respectively. This indicates the receiver noise power, including potential interference and receiver thermal noise.

[0124] (23) B2V channel model

[0125] Considering the small-scale attenuation effect in B2V communication, the channel gain is calculated as follows:

[0126]

[0127] Where g0 and d0 represent the path loss constant and reference distance, respectively. It is the distance from BS to the vehicle.

[0128] The achievable content transfer rate from BS b to vehicle v is expressed as:

[0129]

[0130] Among them B b,v and P b,v These represent the communication bandwidth and transmission power between the BS and the vehicle, respectively. This indicates the noise power of the receiver.

[0131] (3) Establish a dynamic edge caching model, and consider service latency and caching costs;

[0132] Based on the Zipf distribution, we can assume that file popularity follows a certain pattern. Therefore, the popularity of edge caching tasks can be represented as follows:

[0133]

[0134] Here, j represents the Zipf index, indicating the skewness of popularity, and ξ represents the ranking of content request frequency.

[0135] Further, definition and Place vectors for UAV u and RSU r in each content and time slot, where u represents the unmanned aerial vehicle (UAV), f represents the cache task, t represents the time slot, and r represents the roadside unit.

[0136] Where α u,f,t ,β r,f,t ∈{0,1}; In addition, the total size of the cached content in each UAV and RSU should be within the available cache capacity Qu and Qr, respectively, where Qu represents the available cache capacity of the unmanned aerial vehicle and Qr represents the available cache capacity of the roadside unit;

[0137] Based on the above analysis, service latency can be expressed as:

[0138]

[0139] The cost of caching is described as:

[0140]

[0141] Here, ρ u ρ r and ρ b These represent the unit cost of renting cache resources from UAV u, RSU r, and BS b, respectively. u N r and N b S represents the number of edge nodes (i.e., UAV, RSU, and BS) per unit area. u S r and S b . These represent the number of popular content cached on UAV u, RSU r, and BS b, respectively.

[0142] (4) The allocation of cache resources between UAV and RSU is designed as a joint optimization problem to minimize the total cost of edge task delivery. In the air-ground cooperative edge caching architecture, the two indicators of service latency and cache cost are weighed.

[0143] The mathematical formulation of this optimization problem is as follows:

[0144] P1:

[0145] st(a)α u,f,t ,β r,f,t ∈{0,1},

[0146] (b)

[0147] (c)

[0148] (d)

[0149] (e)

[0150] Where ∈{0,1} is a weighted parameter, constraint (a) is a binary variable constraint representing the indivisibility of content; constraint (b) stipulates that a vehicle can only be served by one edge node at a time; constraint (c) limits the maximum amount of content that can be cached in UAVs and RSUs; constraint (d) stipulates the maximum allowed connection time between UAVs, RSUs and vehicles;

[0151] Unlike the traditional knapsack problem, the proposed problem P1 (i.e., MMKP) combines multiple dimensions (i.e., transmission latency and caching capacity) and multiple knapsacks (i.e., multiple edge nodes). Assigning content to different edge nodes consumes varying amounts of communication and caching resources. While existing research has proposed precise algorithms (such as branch and boundary algorithms and dynamic programming) to address this problem, brute-force methods become prohibitively expensive as the problem size increases. Inspired by the KS framework, this invention proposes an ACCO algorithm for large-scale MMKP.

[0152] (5) The ACCO algorithm based on the kernel search heuristic framework and greedy strategy is used to optimize the dynamic edge caching model.

[0153] Kernel search is a heuristic framework for solving multidimensional knapsack and combinatorial choice problems. It selects a set of core variables that may be part of the optimal solution and solves a series of subproblems involving the core set and a selected set of additional variables. This problem-agnostic framework can be applied to a wide range of problems.

[0154] Specifically, to improve the performance of the KS framework, this invention proposes an ACCO algorithm based on unsupervised learning and a greedy strategy. The main steps of the algorithm are described below:

[0155] Initialization phase: Relax the linear programming of the original problem, use the k-means clustering algorithm to classify the content, and obtain an upper bound.

[0156] Iterative solution process: Large-scale MMKP problems are decomposed into a series of subproblems, and then the CPLEX tool is used to iteratively solve them to obtain feasible solutions.

[0157] Dynamic adjustment phase: The dimensions and solution time of the subproblems are dynamically adjusted. Specifically, the solution results are further optimized by using a greedy load balancing strategy.

[0158] Figure 3 The ACCO algorithm process has been demonstrated. The detailed algorithm process will be described below.

[0159] (51) Initialization phase

[0160] A good upper bound is crucial for pruning infeasible search spaces and achieving fast convergence. As shown in P2, by changing the constraints (14a) on the objective function, a linear programming relaxation of the original problem P1 is obtained.

[0161] P2:

[0162] st(a) 0 ≤ α u,f,t ,β r,f,t ≤1,

[0163] (14b)~(14e).

[0164] First, solve for P2. During the continuous relaxation process, select all variables with positive values ​​as the initial kernel set Δ, and then obtain the upper bound z. LP Then, the K-means clustering algorithm is used to eliminate infeasible solutions z. LP Transforming into a feasible solution, this algorithm is a popular prototype-based clustering algorithm. It divides |F| content into |K| groups (i.e., the number of types of edge nodes), where |K| < |F|. Each piece of content belongs to the group closest to the mean (group centroid), which serves as the prototype of the group. The distance between different cluster centroids is the greatest. This invention provides a series of feature data (x1, ..., x...) for the content. F (i.e., service latency, caching cost, and content size) and collections The K-means clustering algorithm minimizes the sum of squares within a cluster, i.e., the square of the Euclidean distance. Formally, its goal is to find:

[0165]

[0166] Where μ k Is the user in cluster s k The average position in.

[0167] (52) Iterative solution process

[0168] At this stage, the KS framework is used to traverse the buckets, which are sequences of subproblems, and obtain feasible solutions.

[0169] First, variables not in the kernel set are divided into bucket sequences BK. i Let NB be the number of buckets generated. This decomposition transforms a large-scale integer linear programming (ILP) problem into a series of smaller subproblems. Therefore, P2 can be transformed into:

[0170] P3:

[0171] st(a)

[0172] (b)

[0173] (14e), (15a).

[0174] Here, m represents the number of items contained in each subproblem, which will be dynamically adjusted in later stages. and It is the adjusted fixed value. and It can be represented as:

[0175]

[0176]

[0177] This led to the construction of Λ∪BK i The set (i.e., the current bucket) is solved using the CPLEX tool by adding the following two constraints.

[0178]

[0179] z cut =z * +1

[0180] Where z cut It is the cutoff value of the objective function of the problem.

[0181] Secondly, states are labeled based on the quality and time required to solve subproblems, which is important for evaluating instance difficulty. If a feasible (but not necessarily optimal) solution is found, the state is labeled feasible. If no feasible solution is found before the time limit, the state is labeled timeLimit. If the outcome of a subproblem becomes infeasible due to the deadline, the state is labeled cutoff. Finally, if a solution to a subproblem is close to optimal before the time limit is reached, the state is labeled optimal. When a state is feasible or optimal, the best feasible solution is determined by x. * This indicates that its value is z. * .

[0182] (53) State adjustment phase

[0183] Based on this, the size of the subproblems and the solution time are dynamically adjusted according to the instance information collected during the iterative solution process. If the percentage of subproblems causing timeouts is less than a predetermined value ι, the instance is considered fast. Otherwise, it is considered slow. At the end of each bucket iteration, the initial bucket size m and the higher computation time limit t are adjusted. op The size Δ of the bucket is updated to fast or slow.

[0184] This invention designs a greedy load balancing strategy to select items to be included in the kernel. This strategy considers service latency, caching costs, and previous request frequency to optimize the caching location for UAVs and RSUs. It assumes that different vehicles may request the same content in the same area. Therefore, the value ratio e is used... f,u and e f,r Defined as:

[0185]

[0186]

[0187] in W is a weighted parameter. f This indicates the request frequency of content f. If e f,u ,e f,r If the value is greater than 1, the vehicle will request a BS (Balanced Shift). If e f,u ,e f,r If e < 1, then these values ​​are compared. f,u <e f,r Content is cached on the User Access Container (UAV) if it is large; otherwise, it is cached on the Receiving Subsequent Unit (RSU). It can be observed that when the content size is large, the impact of service latency outweighs the caching cost. However, for efficiency, frequently requested large content can also be cached on the UAV or RSU instead of the Service Container (BS). Therefore, e f,u and e f,r The ratio can help the KS framework improve the performance of the proposed algorithm.

[0188] Finally, this embodiment uses Matlab to build a simulation model and implement the ACCO algorithm. The simulation scenario includes 5 UAVs and 5 RSUs. The UAVs fly along a fixed trajectory with a radius of [missing information]. The altitude is h, set to 400 meters and 100 meters respectively. Each RSU's antenna covers an area of ​​500 meters around it. The size of each content is in the range of [0, 50] MB. In the simulation, UAVP... u,v RSUP r,v and BSP b,v The transmit power was set to 0.2W, 0.1W, and 1W, respectively. The results show that the allocated bandwidth B...u,v B r,b and B b,v The frequencies are 3MHz, 1MHz, and 2.5MHz, respectively. Due to unpredictable hardware factors, the unit cost of renting cache resources from edge nodes is randomly set within a certain range.

[0189] The comparison algorithms include the optimal solution (OPT) solved using the CPLEX tool, and four existing optimization algorithms, which are shown below:

[0190] The CRS-PSO algorithm: It modifies the Particle Swarm Optimization (PSO) algorithm by adding cache replacement to overcome slow convergence and provide better performance.

[0191] The ISA algorithm improves simulated annealing (SA) by extracting two solution spaces, thereby obtaining a higher quality approximate solution.

[0192] The GABG algorithm combines a genetic algorithm (GA) with a greedy strategy to reduce solution time and improve solution accuracy.

[0193] The PBIG algorithm: It is a simple extension of the iterative greedy algorithm, which involves creating an initial set of solutions and iteratively improving them through destruction and reconstruction phases.

[0194] Figure 3 The service latency and caching cost of six algorithms under different request counts are described. It can be seen that service latency and caching cost increase with the number of requests. Among these algorithms, the ACCO algorithm provides near-optimal performance in solving the optimization problem. This is because MMKP is NP-hard; therefore, CRS-PSO, ISA, and GABG algorithms easily get trapped in local optima, and the PBIG algorithm's parameters are randomly set (i.e., the degree of destruction and the deterministic rate), with its greedy rule only considering the ratio of resource consumption to profit. However, the ACCO algorithm uses the CPLEX tool to iteratively solve subproblems, considering the optimization objective (i.e., service latency and caching cost) and request frequency, placing content with similar characteristics on the most suitable edge nodes.

[0195] The impact of the Zipf parameter j of the popularity distribution on service latency and caching costs is detailed below. Figure 4Generally, the larger the Zipf parameter j, the more concentrated the user's preferences are on highly popular content. However, the ACCO algorithm needs to consider both service latency and caching costs simultaneously, so the results do not show a monotonic increase or decrease. Specifically, when j > 0.6, the weighted sum of service latency and caching costs increases steadily. The ACCO algorithm has the highest caching efficiency when j = 0.6. Furthermore, changing the value of j has the greatest impact on the PBIG algorithm.

[0196] I am Figure 5 The table shows the percentage of service latency and caching cost for each algorithm across 1000 requests. The first observation is that the caching costs of the six algorithms are not significantly different because the content sizes are similar. This suggests that only when the content is large enough will the BS alleviate the pressure of high caching costs with less service latency. The second observation is that the ACCO algorithm can significantly reduce caching costs. This is because it combines a k-means clustering algorithm with a greedy load balancing strategy to achieve optimal cost performance between content and edge nodes.

[0197] Furthermore, in the following Figure 6 The figure compares the runtime of each algorithm for a certain number of requests. As shown in the figure, the runtime of all six algorithms increases with the number of requests. While the CPLEX tool can provide the optimal solution, it inevitably requires significant computational overhead in large-scale scenarios. As mentioned above, the ACCO algorithm uses the KS framework to transform a large-scale problem into a series of subproblems, which effectively reduces computational overhead. The PBIG algorithm reduces the dimensionality of the problem by breaking and building up partial solutions. In summary, the results show that the ACCO algorithm has advantages under different settings, namely, popularity parameters, number of requests, and runtime.

Claims

1. An adaptive cooperative air-to-ground edge caching method for dynamic vehicle-to-everything (V2X) scenarios, characterized in that, Includes the following steps: (1) Construct a dynamic edge caching system to dynamically network vehicles, unmanned aerial vehicles (UAVs), roadside units (RSUs), and base stations (BSs) that have edge computing task caching capabilities; (2) Establish channel models between UAV, RSU, BS and vehicles, including U2V channel model, R2V channel model and B2V channel model, and calculate the corresponding transmission rates as follows: The transmission rate between the unmanned aerial vehicle u and the vehicle v is calculated as follows: in and These represent the bandwidth and transmission power between UAV u and vehicle v, respectively. Indicates noise power. This represents the average path loss. The transmission rate from roadside unit r to vehicle v is expressed as: in and These represent the communication bandwidth and transmit power between the RSU and the vehicle, respectively. This represents the receiver noise power, including potential interference and receiver thermal noise. The channel gain between the second edge node and the vehicle; The achievable content transfer rate from base station BS b to vehicle v is expressed as follows: in and These represent the communication bandwidth and transmission power between the BS and the vehicle, respectively. Indicates the noise power of the receiver. For the channel gain between the central server node and the vehicle; (3) Establish a dynamic edge caching model, taking into account service latency and caching costs; definition = and = For each content and time slot, place vectors for UAV u and RSU r in the cache, where u represents the unmanned aerial vehicle (UAV). Indicates a cache task. The time slot is represented by r, which represents the roadside unit. in ; In addition, the total size of cached content in each UAV and RSU should be set within the available cache capacity. and Inside, This indicates the available cache capacity of the unmanned aerial vehicle. Indicates the available cache capacity of the roadside unit; (4) The allocation of cache resources between UAV and RSU is designed as a joint optimization problem to minimize the total cost of edge task delivery. In the air-ground cooperative edge caching architecture, the two indicators of service latency and cache cost are weighed. This optimization problem is limited to a multi-dimensional knapsack problem, and its mathematical formulation is as follows: s.t. (a) , (b) , (c) , (d) , (e) , in It is a weighted parameter. Due to service delay; For caching costs; For caching tasks The size of the content; constraint (a) is a binary variable constraint, representing the indivisibility of the content; constraint (b) stipulates that a vehicle can only be served by one edge node at a time; constraint (c) limits the maximum amount of content that can be cached in UAVs and RSUs; constraint (d) specifies the maximum allowed connection time between UAVs, RSUs and vehicles; (5) The ACCO algorithm based on the kernel search heuristic framework and greedy strategy is used to optimize the dynamic edge caching model. The solution of the dynamic edge caching model includes the following three stages: Initialization phase: Relax the linear programming of the original problem, use the k-means clustering algorithm to classify the content, and obtain an upper bound; Iterative solution process: The joint optimization problem of cache resource allocation between UAV and RSU is decomposed into a series of sub-problems, and then the CPLEX tool is used to iteratively solve the problem to obtain a feasible solution; Dynamic adjustment phase: The dimensions and solution time of the subproblems are dynamically adjusted, specifically by using a greedy load balancing strategy to further optimize the solution results.

2. The adaptive cooperative air-to-ground edge caching method in a dynamic vehicle-to-everything (V2X) scenario according to claim 1, characterized in that, In step (1), the UAV flies over the area covered by the vehicle connectivity system to provide flexible connectivity, and the RSU is deployed on the road to provide stable connectivity. Each UAV and RSU is connected to an edge caching server; BS stands for cloud server, which contains all requested content; The UAV, RSU, and BS communicate with the vehicle via a cellular vehicle-to-everything (V2X) interface.

3. The adaptive cooperative air-to-ground edge caching method in dynamic vehicle-to-everything (V2X) scenarios according to claim 2, characterized in that, In a dynamic edge caching system, the unmanned aerial vehicle U, roadside unit R, base station B, and vehicle V are represented as U = {1, 2, ..., U}, R = {1, 2, ..., R}, B = {1, 2, ..., B}, and V = {1, 2, ..., V}. Due to the high dynamism, let Lu and Lr represent the finite contact time between the UAV u∈U, RSU r∈R, and the vehicle v∈V, respectively; BS b∈B provides a stable link for the vehicle, and let the set F={1,2,...,F} represent the index set of available capacity, caching the task. The size is defined as The capacity vector formed by the sizes of all cached tasks is . In addition, the entire timeline is divided into T time periods of equal length, with each time period indexed by t.

4. The adaptive cooperative air-to-ground edge caching method in a dynamic vehicle-to-everything (V2X) scenario according to claim 1, characterized in that, Step (2) for calculating the edge task transmission rate under the U2V channel model, R2V channel model, and B2V channel model includes the following considerations: For the U2V channel model, considering both line-of-sight and non-line-of-sight aspects of the U2V channel, its expression is as follows: in and This represents the carrier frequency and the distance from UAV u to vehicle v; and The shading and diffraction losses are for LoS and NLoS, respectively. and These represent the line-of-sight and non-line-of-sight path loss correction coefficients, respectively. For UAVs and vehicles, the probabilities of LoS and NLoS links depend on their altitude and environment, and can be expressed as: Where a and b are constants related to the environment. The angle of elevation is expressed in radians. and These represent the probabilities of Loss of Sight (LOS) and non-LOS links, respectively. Therefore, the average path loss associated with LoS and NLoS It is expressed as follows: For the R2V channel model, assuming the time-varying attenuation channel follows a Rician distribution, the expression for calculating the channel gain is as follows: in It is a constant. and These represent the antenna gain and path loss exponents, respectively. For carrier frequency, Indicates from RSU r to vehicle The distance; For the B2V channel model, considering the small-scale attenuation effect in B2V communication, the channel gain is calculated as follows: in and These represent the path loss constant and the reference distance, respectively. It is the distance from BS to the vehicle.

5. The adaptive cooperative air-to-ground edge caching method in a dynamic vehicle-to-everything (V2X) scenario according to claim 1, characterized in that, In step (3), for the dynamic edge caching model, according to the Zipf distribution, for the edge caching task... Popularity is expressed as follows: Representing the Zipf index, used to characterize the skewness of popularity. This indicates the ranking based on the frequency of content requests.

6. The adaptive cooperative air-to-ground edge caching method in a dynamic vehicle-to-everything (V2X) scenario according to claim 1, characterized in that, The service latency of the dynamic edge caching model described in step (3) is represented as follows: The cost of caching is described as: , and These represent the unit cost of renting cache resources from UAV u, RSU r, and BS b, respectively. , and This represents the number of edge nodes per unit area, namely UAVs, RSUs, and BSs. , and These represent the number of edge cache tasks cached on UAV u, RSU r, and BS b, respectively.

7. The adaptive cooperative air-to-ground edge caching method in a dynamic vehicle-to-everything (V2X) scenario according to claim 6, characterized in that, The initialization phase includes optimizing the constraints (a) of the objective function to obtain the linear programming relaxation of the original problem P1, which is then transformed into solving problem P2. The mathematical representation of problem P2 is as follows: For problem P2, the K-means clustering algorithm minimizes the sum of squares within each cluster, i.e., the square of the Euclidean distance, and the objective expression is as follows: in It is the user in the cluster The average position in For caching tasks Feature data vector The iterative solution process involves transforming problem P2 into problem P3, mathematically represented as follows: P3: s.t. (a) , (b) , In the formula, m represents the number of contents contained in each subproblem, which will be dynamically adjusted in later stages; , , , and It is an adjusted fixed value. and Represented as: Therefore, it is constructed The set can be solved using the CPLEX tool by adding the following two constraints. in It is the cutoff value of the objective function of the problem. The optimal solution is the optimal objective function value found in the current iteration. The dynamic adjustment phase includes designing a greedy load balancing strategy to select items to include in the kernel; assuming different vehicles might request the same content in the same region, the value ratio involved in this strategy is adjusted accordingly. and Defined as: in It is a weighted parameter. Indicates cache task The request frequency, if , >1, the vehicle will request BS; if , If <1, then these values ​​are compared; if < If the content is cached on the UAV, then the content is cached on the RSU.

Images