Deployment optimization and computation offloading method for space-ground integrated mobile edge computing system

Abstract

The present invention provides a method for optimizing the deployment and offloading computation of an integrated space-air-ground mobile edge computing system, including a ground layer of M GTs, an air layer of N UAVs, and a space layer composed of one LEO satellite; the sets of GTs and UAVs are represented as M={1,2,…,M} and N={1,2,…,N}, respectively; in the ground layer, GTs with limited computing capabilities are distributed in areas lacking communication infrastructure and perform tasks with computing requirements; UAVs in the air layer serve as edge nodes to provide edge computing services for GTs; in the space layer, LEO satellites provide centralized cloud computing services for all GTs in their coverage area; and GTs are allowed to offload part of their computing tasks to UAVs and LEO satellites for execution. The application of this technical solution can minimize the average response time of each computing task in the system by optimizing the deployment location of UAVs and the offloading strategy of computing tasks.

Description

Technical Field

[0001] The present invention relates to the field of space-ground-integrated computing technology, and in particular to a method for deployment optimization and computing offloading of a space-ground-integrated mobile edge computing system. Background Art

[0002] The rapid development of the Internet of Things and related science and technology has led to the emergence of a variety of application services, which has brought great convenience to people's study, life and work. However, emerging computing-intensive applications are usually highly sensitive to latency, so they place high demands on the computing power of terminal devices during execution. Due to the limitations of existing hardware technology levels and portability, the limited storage and computing resources of terminal devices are difficult to independently meet the needs of the above applications, which greatly reduces the performance of computing tasks when executed on terminal devices.

[0003] Mobile edge computing (MEC) technology deploys servers at the edge of the network and is considered to be a forward-looking technical means to address the above challenges. In mobile edge computing technology, terminal devices can offload computing tasks to be executed to nearby MEC servers, which can overcome the network congestion problem caused by centralized processing in traditional mobile cloud computing, and has a shorter transmission distance, less transmission time and energy consumption, thereby effectively improving the user's service quality. However, the location of MEC servers is usually fixed and cannot be flexibly changed according to the needs of mobile users, which limits the capabilities of MEC systems.

[0004] In recent years, with the continuous development of unmanned aerial vehicle (UAV) technology, UAV has received widespread attention in the field of wireless communications. UAV has been used in some scenarios where communication infrastructure is limited, such as disaster response and battlefield communications. On this basis, the air-space-ground integrated cloud-edge hybrid MEC system has begun to attract academic attention. Generally speaking, the air-space-ground integrated system consists of three parts: the space layer network, the air layer network, and the ground network. The air layer network consists of UAVs equipped with MEC servers, which can provide remote wireless access and low-latency edge computing services for ground terminal devices (GT) in a certain area. The low Earth orbit (LEO) satellites in the space layer network can cover the entire area and provide ubiquitous cloud computing services for all ground devices. At present, some works have studied the computing offloading mechanism enabled by space satellites. However, these works mainly focus on air-ground cooperation and do not consider air-space-ground coordination. In addition, although some works have studied the key technologies of air-space-ground integrated MEC systems, most of these works only consider the scenarios supporting a single UAV and cannot be applied to actual complex scenarios. In addition, in order to simplify the model, most of the works only adopt a binary offloading strategy, which will lead to low resource utilization of various terminals in the system. Finally, existing work on UAV-enabled MEC systems usually aims to optimize system energy consumption, energy efficiency or throughput, and the research on latency, especially the overall average response latency of the system is still slightly insufficient. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a deployment optimization and computing offloading method for an integrated air-space-ground mobile edge computing system, which minimizes the average response time of each computing task in the system by optimizing the deployment position of UAVs and the offloading strategy of computing tasks.

[0006] To achieve the above objectives, the present invention adopts the following technical scheme: a method for optimizing the deployment and computing offloading of an integrated air-ground-space mobile edge computing system, including a ground layer of M GTs, an aerial layer of N UAVs, and a space layer consisting of 1 LEO satellite; the sets of GTs and UAVs are represented as M = {1, 2, ..., M} and N = {1, 2, ..., N}, respectively; in the ground layer, GTs with limited computing capabilities are distributed in areas lacking communication infrastructure and perform tasks with computing requirements; UAVs in the aerial layer serve as edge nodes to provide edge computing services for GTs; in the space layer, LEO satellites provide centralized cloud computing services for all GTs in their coverage area; and GTs are allowed to offload part of their computing tasks to UAVs and LEO satellites for execution.

[0007] In a preferred embodiment, a three-dimensional Euclidean coordinate system is used to represent the positions of various ground terminal devices and UAVs. The position of the i-th GT in the current time slot is represented by the three-dimensional coordinate system Indicates that, accordingly, the current position of the jth UAV is expressed in three-dimensional coordinates Where H represents the flight altitude of UAVs;

[0008] Take a 4-tuple I i = <D i ,S i ,O i ,E i > represents the computing task that the i-th GT needs to perform in the current time slot, where D i Represents computing task I i The input data size; S i Indicates the number of computing resources required to execute the unit computing task input; i Indicates the size of the task execution result; E i Indicates the maximum tolerable delay of the computing task; the computing task execution result O i The size of is much smaller than the input of the task D i , computing task I i The completion delay shall not exceed the maximum tolerable delay E i ;

[0009] The calculation task I i The partial offloading strategy is adopted to allow GTs to divide their computing tasks and offload part of the computing tasks to UAVs and LEO satellites for execution. The proportion of the local execution part and the subtasks offloaded to UAVs and LEO satellites are respectively expressed as and To express;

[0010]

[0011] When the i-th GT is not within the coverage of any UAV, the GT cannot offload part of the computing tasks to the UAVs for execution.

[0012] In a preferred embodiment, the execution process of the original computing task is divided into three different modes according to the different execution locations of each subtask, namely: a local computing mode executed locally, an edge computing mode executed on the UAV, and a cloud computing mode executed on the LEO satellite; these three computing modes are executed in parallel;

[0013] 1) In local computing mode, define parameters is the computing frequency of the i-th GT local CPU, according to the computing task I i And local uninstall ratio The definition of the local task size to be processed is Therefore, the delay caused by local calculation is expressed as

[0014]

[0015] 2) In edge computing mode, the latency of executing computing tasks will consist of three parts, namely: the transmission latency of offloading the task to the associated UAV, the computing latency of executing on the UAV, and the backhaul latency of returning the computing results to the GT;

[0016] The wireless channel between the i-th GT and the j-th UAV is modeled as a line-of-sight channel link model; the wireless channel power gain between the i-th GT and the j-th UAV is quantified as

[0017]

[0018] Where h0 represents the channel gain in the wireless channel when the reference distance d0 = 1 meter, d ij It represents the distance between the i-th GT and the j-th UAV, which can be expressed as

[0019]

[0020] Therefore, the data transmission rate between the i-th GT and the j-th UAV is further expressed as

[0021]

[0022] Where B represents the system bandwidth between GTs and UAVs, P i represents the transmission power of the i-th GT; σ 2 Represents the background noise power in the channel.

[0023] Define a binary variable β ij represents the relationship between GTs and UAVs, where β ij =1 means that the i-th GT decides to offload part of the task to the j-th UAV for execution, otherwise β ij = 0; it is stipulated that each GT can at most offload part of the task to one UAV for execution, so there are the following constraints:

[0024]

[0025] In addition, the upper limit of the number of concurrent operations per UAV in the current time slot is specified, so there are the following constraints:

[0026]

[0027] For the i-th GT, its task completion delay in edge computing mode consists of two parts: transmission delay and calculation delay, which can be specifically expressed as

[0028]

[0029] in, It represents the CPU computing frequency assigned to the associated GT by the jth UAV;

[0030] 3) In the cloud computing mode, GTs and LEO satellites will use direct ground-to-space transmission technology to offload some computing tasks to remote execution, and the data transmission rate between all GTs and LEO satellites is uniformly defined as R s , define the parameters represents the CPU computing frequency allocated to each GT by the LEO satellite. Therefore, the task completion delay in the cloud computing mode is expressed as

[0031]

[0032] Since the local computing mode, edge computing mode, and cloud computing mode can be executed in parallel, for the i-th GT, the completion delay of its computing task is T i Should be expressed as

[0033]

[0034] The system average response delay is defined as the average of the completion delays of all computing tasks in the system, which is expressed as follows:

[0035]

[0036] The optimization problem of joint deployment and computation offloading is formulated as:

[0037]

[0038] Among them, C1 means that the completion delay of each task cannot exceed the maximum tolerable delay of the task; C2 and C3 represent the constraints on the task offloading rate; C4-C6 represent the association constraints between GTs and UAVs; C7 means that each UAV can only provide offloading services for GTs within a limited range.

[0039] In a preferred embodiment, the following encoding mechanism is adopted: the population is composed of a number of particles, each particle represents a deployment plan of a UAVs cluster; each particle is encoded as a set of N two-dimensional vectors, each of which represents the horizontal coordinates of a UAV in the current time slot; specifically, in the tth iteration, the kth particle in the population is represented as

[0040]

[0041] in It represents the horizontal coordinate of the jth UAV in the t-round iteration, which is specifically expressed as

[0042]

[0043] in and They represent the horizontal and vertical coordinates of the deployment position of the jth UAV in the kth particle, respectively;

[0044] The PSO-GA algorithm improves the particle update process of the traditional PSO algorithm by introducing the crossover and mutation operators in the GA algorithm. At this time, the update method of the kth particle is changed to

[0045]

[0046] Among them, M u () represents the mutation operator of the PSO-GA algorithm, C p () and C g () represents the crossover operator;

[0047] The inertia part of the particle is updated as follows:

[0048]

[0049] Here, r1 is a random value in the interval [0,1]. Indicates randomly selected particles A certain quantile on , and perform random mutation within the threshold range;

[0050] In the PSO-GA algorithm, the traditional individual learning part is combined with the social learning part and the crossover operator of the GA algorithm. The corresponding update methods are:

[0051]

[0052]

[0053] Among them, r2 and r3 are two random values ​​in the interval [0,1]; the crossover operator randomly selects two quantiles on the particle and uses pBestk or gBest Replace the values ​​between the corresponding quantiles.

[0054] A linear inertia weight factor adjustment strategy is adopted, which will make the inertia weight factor w decrease linearly with the increase of the number of iterations. The specific adjustment strategy is as follows:

[0055]

[0056] In the formula, w max and w min They are the maximum and minimum values ​​set by w, iters max Indicates the agreed maximum number of iterations, iters cur Indicates the current iteration number;

[0057] A similar linear adjustment strategy is used as follows:

[0058]

[0059]

[0060] Among them, c 1_start and c 2_start They represent the initial values ​​of parameters c1 and c2 before the iteration starts, respectively. 1_end and c 2_end Respectively represent the final values ​​of parameters c1 and c2 in the iteration.

[0061] In a preferred embodiment, the UAVs deployment optimization algorithm based on PSO-GA is as follows:

[0062] Step 1: Randomly initialize the population, including the population size pN and the maximum number of iterations iter max And particle U i , and initialize the local optimal solution pBest of each population i , and the global optimal solution gBest;

[0063] Step 2: When the algorithm execution conditions are met, the particle population is updated through the crossover and mutation operations of the particles, and Algorithm 2 is called to obtain the computational offloading scheme. The fitness of each particle is calculated according to formula (11), and the local optimal solution pBest is updated according to the fitness. i And the global optimal solution gBest;

[0064] Step 3: Output the final global optimal solution gBest.

[0065] In a preferred embodiment, the proposed PSO-GA algorithm lower layer needs to achieve computational offloading optimization under the given UAVs deployment scheme, including user association { α} and uninstall rate {β} optimization, as follows:

[0066]

[0067] All GTs are divided into three types according to their distribution:

[0068] 1) The GT is not within the coverage of any UAV;

[0069] 2) The GT is within the coverage of only one UAV;

[0070] 3) The GT is located within the coverage area of ​​multiple UAVs.

[0071] Subsequently, we give the priorities of the above three categories in descending order; the first category of GTs has the highest priority, because this type of GTs cannot be associated with UAVs and does not occupy the computing resources of UAVs, so it will not affect the offloading strategies of the other two categories of GTs; the second category of GTs has the second level of priority, because this type of GTs can only be associated with the nearest UAV; the third category of GTs selects the appropriate UAV for association among multiple UAVs, so it has the lowest priority.

[0072] Then, we decide the task offloading rate of GTs; for the first type of GTs mentioned above, since they cannot be associated with UAVs, that is, At this time, the original computing task is only divided into the local execution part and the LEO satellite execution part; for the second and third types of GTs, the original task is divided into the local execution part, the UAV execution part and the LEO satellite execution part.

[0073] In a preferred embodiment, the calculation offloading optimization algorithm based on Greedy is as follows:

[0074] 1) Divide the GTs into three sets according to the position between GTs and UAVs;

[0075] 2) For the GTs in the first set, the unloading rate is directly calculated. For the GTs in the second and third sets, the association relationship with the UAVs is first determined, and then the unloading rate is calculated;

[0076] 3) Output the final correlation set α and uninstall rate set β.

[0077] Compared with the prior art, the present invention has the following beneficial effects: In the present invention, a joint optimization algorithm PG-G for UAVs deployment and computing task offloading in an air-space-ground integrated MEC system based on PSO-GA and Greedy algorithm is proposed. The algorithm can converge to a stable value in a finite number of iterations and has a better optimization effect than the benchmark algorithm. The present invention is applicable to an air-space-ground integrated MEC system with multiple UAVs and multiple ground devices. By optimizing the deployment position of UAVs and the offloading strategy of computing tasks, the average response time of each computing task in the system is minimized. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1A schematic diagram of an air-ground integrated MEC system according to a preferred embodiment of the present invention;

[0079] Figure 2 A schematic diagram of an example of particle encoding according to a preferred embodiment of the present invention;

[0080] Figure 3 A schematic diagram of a variation operation of a preferred embodiment of the present invention;

[0081] Figure 4 A schematic diagram of a crossover operation according to a preferred embodiment of the present invention;

[0082] Figure 5 It is a PG-G iterative curve of a preferred embodiment of the present invention;

[0083] Figure 6 The average system response delay of different algorithms when Rmax=40m in the preferred embodiment of the present invention;

[0084] Figure 7 The average system response delay of different algorithms when Rmax=80m in the preferred embodiment of the present invention;

[0085] Figure 8 The average system response delay of different algorithms when Rmax=40m in the preferred embodiment of the present invention;

[0086] Fig. 9 The average system response time of different algorithms when Rmax=80m in the preferred embodiment of the present invention. DETAILED DESCRIPTION

[0087] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0088] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0089] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0090] A deployment optimization and computation offloading method for an air-ground integrated mobile edge computing system, such as Figure 1The air-space-ground integrated MEC system shown in the figure is composed of a three-layer network architecture, namely: a ground layer with M GTs, an aerial layer with N UAVs, and a space layer composed of 1 LEO satellite. For ease of analysis, the present invention represents the set of GTs and UAVs as M = {1, 2, ..., M} and N = {1, 2, ..., N}, respectively. At the ground layer, GTs with limited computing power are distributed in areas where communication infrastructure is lacking (such as mountainous areas, oceans, and disaster-stricken areas), and perform some tasks with certain computing requirements. In order to alleviate the pressure of GTs to perform computing tasks, considering the lack of coverage of ground cellular networks in this area, UAVs in the aerial layer can serve as edge nodes to provide edge computing services for GTs. In addition, in the space layer, LEO satellites can provide centralized cloud computing services for all GTs in their coverage area. Although GTs can perform their computing tasks locally, due to the limited computing power of GTs, they may not be able to complete their tasks independently. Therefore, GTs are allowed to offload part of their computing tasks to UAVs and LEO satellites for execution.

[0091] The present invention uses a three-dimensional Euclidean coordinate system to represent the positions of various ground terminal devices and UAVs. The position of the i-th GT in the current time slot can be expressed by the three-dimensional coordinate system Indicates that, accordingly, the current position of the jth UAV can be expressed in three-dimensional coordinates Where H represents the flight altitude of UAVs. It should be noted that, in order to simplify the model, the present invention assumes that all UAVs are deployed at the same altitude in the current time slot. In addition, since the present invention studies the system performance in a quasi-static scenario, this means that the positions of all UAVs and GTs will remain unchanged in the considered time slot. In the present invention, a 4-tuple I is used. i = <D i ,S i ,O i ,E i > represents the computing task that the i-th GT needs to perform in the current time slot, where D i Represents computing task I i The input data size (in bits); S i Indicates the number of computing resources required to execute the unit computing task input (in cycles / bit); i Indicates the size of the task execution result (in bits); E i Indicates the maximum tolerable delay of the computing task (in seconds). In this definition, the computing task execution result O i The size of is usually much smaller than the input size of the task D i , so it can be ignored. In addition, the calculation task I iThe completion delay shall not exceed the maximum tolerable delay E i , because there is no point in completing a task after its deadline.

[0092] Due to the limited computing power of GTs, it may not be possible to independently i The calculation task needs to be completed within 1 second, so the calculation task I i In order to make full use of the computing resources of the devices at the ground layer, air layer and space layer, the present invention adopts a partial offloading strategy, allowing GTs to divide their computing tasks and offload part of the computing tasks to UAVs and LEO satellites for execution. The proportion of the local execution part and the subtasks offloaded to UAVs and LEO satellites are respectively expressed as and Similar to other existing works, this paper assumes that the partial offloading strategy will not generate additional computing tasks, so

[0093]

[0094] It should be noted that when the i-th GT is not within the coverage of any UAV, the GT cannot offload part of the computing tasks to the UAVs for execution.

[0095] According to the different execution locations of each subtask, the present invention divides the execution process of the original computing task into three different modes, namely: the local computing mode executed locally, the edge computing mode executed on the UAV, and the cloud computing mode executed on the LEO satellite. Since the computing resources consumed by the three computing modes belong to different terminals, the three computing modes can be executed in parallel.

[0096] 1) In local computing mode, define parameters is the computing frequency of the i-th GT local CPU, according to the computing task I i And local uninstall ratio The definition of can be obtained that the size of the local task to be processed is Therefore, the delay caused by local calculation can be expressed as

[0097]

[0098] 2) In edge computing mode, the latency of executing computing tasks will consist of three parts: the transmission latency of offloading the task to the associated UAV, the computing latency of executing on the UAV, and the backhaul latency of returning the computing result to the GT. Since the size of the computing result is much smaller than the input of the computing task, the backhaul latency can usually be ignored.

[0099] The present invention models the wireless channel between the i-th GT and the j-th UAV as a line-of-sight (LOS) link model. Therefore, the wireless channel power gain between the i-th GT and the j-th UAV is quantified as:

[0100]

[0101] Where h0 represents the channel gain in the wireless channel when the reference distance d0 = 1 meter, d ij represents the distance between the i-th GT and the j-th UAV, which can be expressed using the Euclidean coordinate system as

[0102]

[0103] Therefore, the data transmission rate between the i-th GT and the j-th UAV can be further expressed as

[0104]

[0105] Where B represents the system bandwidth between GTs and UAVs. Similar to the literature

[32] , the present invention assumes that all UAVs provide all GTs with equal bandwidth allocation services through frequency division multiple access (FDMA); P i represents the transmission power of the i-th GT; σ 2 Represents the background noise power in the channel.

[0106] The present invention defines a binary variable β ij represents the relationship between GTs and UAVs, where β ij =1 means that the i-th GT decides to offload part of the task to the j-th UAV for execution, otherwise β ij = 0. In order to avoid destroying the integrity of the original task by over-dividing the computing task, the present invention stipulates that each GT can at most offload part of the task to one UAV for execution, so there are the following constraints:

[0107]

[0108] In addition, since the concurrent capability of UAVs is limited, it may not be possible to respond to too many offloading requests at the same time. In order to avoid network congestion and queuing delay caused by too many requests, the present invention stipulates the upper limit of the concurrent number of each UAV in the current time slot, so there are the following constraints:

[0109]

[0110] Based on the above analysis, for the i-th GT, its task completion delay in edge computing mode consists of two parts: transmission delay and calculation delay, which can be specifically expressed as

[0111]

[0112] in, It represents the CPU computing frequency allocated by the j-th UAV to the associated GT.

[0113] 3) In the cloud computing mode, GTs and LEO satellites will use direct ground-to-space transmission technology to offload some computing tasks to remote execution. The feasibility of this direct ground-to-space satellite transmission technology has been demonstrated in existing work. Similar to related research, the data transmission rate between all GTs and LEO satellites in this paper is uniformly defined as R s , and due to the long distance between GTs and LEO satellites, this value is usually smaller than the transmission rate between GTs and UAVs. In addition, the definition parameter represents the CPU computing frequency allocated to each GT by the LEO satellite. Therefore, the task completion delay in the cloud computing mode can be expressed as

[0114]

[0115] Based on the above analysis, since the local computing mode, edge computing mode and cloud computing mode can be executed in parallel, for the i-th GT, the completion delay of its computing task is T i Should be expressed as

[0116]

[0117] The system average response delay is defined as the average of the completion delays of all computing tasks in the system, which is expressed as follows:

[0118]

[0119] Considering that the system contains multiple GTs, multiple UAVs and a LEO satellite, it is necessary to jointly optimize the deployment of UAVs {u UAV} and task scheduling, including the association relationship between GTs and UAVs {α} and the task offloading rate {β}, to minimize the average response delay of the system. The optimization problem of joint deployment and computation offloading can be expressed as:

[0120]

[0121] Among them, C1 means that the completion delay of each task cannot exceed the maximum tolerable delay of the task; C2 and C3 represent the constraints on the task offloading rate; C4-C6 represent the association constraints between GTs and UAVs; C7 means that each UAV can only provide offloading services for GTs within a limited range.

[0122] 1. The process or method of using a product.

[0123] Since the formally defined joint optimization problem is a mixed integer nonlinear programming problem, the objective function is a non-convex function, and the constraints C4-C7 are non-convex constraints, which makes it difficult to solve the problem directly. The particle swarm optimization algorithm (PSO) does not use the gradient information of the objective function and has no requirements on the continuity and differentiability of the objective function, and has the potential to solve the above optimization problem. However, the traditional PSO algorithm has the following problems in the solution process:

[0124] 1) The optimization problem involves continuous variables ( u UAV , α ) and discrete variables β, which is a typical mixed decision variable optimization problem. The traditional PSO algorithm is difficult to be directly used to solve this type of problem.

[0125] 2) There is a tight coupling between the deployment of UAVs and the scheduling of tasks. On the one hand, the association between GTs and UAVs depends on the deployment of UAVs, because UAVs can only provide offloading services to GTs within their respective coverage areas; on the other hand, in order to obtain the best system performance, the deployment scheme of UAVs needs to be adjusted according to the corresponding computation offloading scheme.

[0126] 3) Traditional PSO algorithms are prone to fall into local optimality and may not be able to obtain optimal or near-optimal solutions.

[0127] Based on the above reasons, it is inefficient to directly use the traditional PSO algorithm to solve the problem. In order to address the above problems, the present invention proposes a PG-G double-layer nested joint optimization method by integrating the PSO-GA algorithm and the Greedy algorithm. In each iteration, the outer layer of the method optimizes the UAVs deployment position through the PSO-GA algorithm, while the inner layer of the method optimizes the computational offloading through the Greedy algorithm.

[0128] 4.1 Outer layer optimization algorithm based on PSO-GA

[0129] The PSO-GA algorithm improves the particle update strategy of the original algorithm by introducing the crossover and mutation operators of GA. This algorithm inherits the advantages of the PSO algorithm such as easy implementation and rapid convergence, while combining the characteristics of GA with good global search capabilities. It can effectively overcome the defect of the traditional PSO algorithm that it is easy to fall into the local optimal solution and obtain a better solution.

[0130] The present invention adopts the following encoding mechanism: the population is composed of several particles, each particle represents a deployment plan of a UAVs cluster; each particle is encoded as a set of N two-dimensional vectors, each of which represents the horizontal coordinate of a UAV in the current time slot. Specifically, in the tth iteration, the kth particle in the population can be represented as

[0131]

[0132] in It represents the horizontal coordinate of the jth UAV in the t-round iteration, which is specifically expressed as

[0133]

[0134] in and They represent the horizontal and vertical coordinates of the deployment position of the j-th UAV in the k-th particle, respectively.

[0135] Figure 2 An example of particle encoding is shown. The particle consists of 6 quantiles, which represent the deployment positions of 6 UAVs. Each quantile consists of a two-dimensional vector, which represents the horizontal coordinate of a UAV. For example, the encoding of quantile 1 is (123, 121), which means that the deployment position of UAV 1 is (123, 121, H).

[0136] The fitness function is used to evaluate the quality of the solution. The present invention uses the system average response delay function defined in formula (11) as the fitness function of the PSO-GA algorithm. Therefore, by substituting the UAVs deployment plan and the computational offloading plan obtained by the Greedy algorithm into formula (11), the fitness of a particle can be obtained. The Greedy algorithm will be introduced in detail in the next section. Since the optimization goal of the present invention is to minimize the average task response delay of the system, the particle with a smaller fitness function value exhibits better performance.

[0137] In the traditional PSO algorithm, each particle moves in a certain direction and speed throughout the problem space. The particles are determined by their own position and speed. They will continuously iterate and update their positions and speeds in the problem search space based on the surrounding particles and their own experience. In this process, the particle speed and position are updated as follows

[0138]

[0139]

[0140] in, and They represent the velocity and position of the kth particle at the tth iteration, pBest k and gBest respectively represent the historical optimal position of the kth particle and the historical optimal position of the entire population after t rounds of iterations. w is the inertia factor, which determines the impact of the speed of the previous iteration on the movement speed of the current generation; c1 and c2 are called individual learning factors and social learning factors, respectively, reflecting the ability to learn the historical optimal value of itself and the historical optimal value of the population. r1 and r2 are two random values ​​in the interval [0,1], which add randomness to the iterative search process.

[0141] The PSO-GA algorithm improves the particle update process of the traditional PSO algorithm by introducing the crossover and mutation operators in the GA algorithm on the basis of the PSO algorithm. At this time, the update method of the kth particle is changed to

[0142]

[0143] Among them, M u () represents the mutation operator of the PSO-GA algorithm, C p () and C g () represents the crossover operator.

[0144] In the above PSO-GA algorithm, the inertia part of the traditional PSO algorithm will be combined with the mutation idea of ​​the GA algorithm, and the inertia part of the particle is updated as follows:

[0145]

[0146] Here, r1 is a random value in the interval [0,1]. Indicates randomly selected particles , and perform random mutation within the threshold range. Figure 3 The result of executing the mutation operator on the particle is shown. In this process, the mutation operator randomly selects a quantile mp1 and randomly changes the encoding on the mp1 bit from (72, 27) to (11, 162).

[0147] In the PSO-GA algorithm, the traditional individual learning part is combined with the social learning part and the crossover operator of the GA algorithm. The corresponding update methods are:

[0148]

[0149]

[0150] Among them, r2 and r3 are two random values ​​in the interval [0,1]. The crossover operator randomly selects two quantiles on the particle and uses pBest k Or replace the values ​​between the corresponding quantiles on gBest.

[0151] Figure 4 The result of performing a crossover operator on a particle is shown in Figure 1. In this process, the crossover operator randomly selects two quantiles cp1 and cp2 on the old particle and uses pBest to replace them. k Or (gBest) The three two-dimensional vectors between the corresponding quantiles, namely (11, 162), (43, 26) and (169, 24), are replaced.

[0152] In the traditional PSO algorithm, the inertia weight factor w in formula (17) can determine the convergence speed and search ability of the algorithm. When the value of w is large, the particles in the PSO-GA algorithm mutate with a greater probability, so the algorithm has a stronger global search ability; when the value of w is small, the probability of mutation of the PSO-GA algorithm is small, and the algorithm has a better local search ability. In the early stage of algorithm execution, more attention is paid to the diversity of problem space search, and as the search deepens, more attention should be paid to the local search ability in the later stage. Therefore, unlike the use of a fixed inertia weight factor in the traditional PSO algorithm, the present invention adopts a linear inertia weight factor adjustment strategy, which will make the inertia weight factor w decrease linearly with the increase of the number of iterations. The specific adjustment strategy is as follows:

[0153]

[0154] In the formula, w max and w min They are the maximum and minimum values ​​set by w, iters max Indicates the agreed maximum number of iterations, iters cur Indicates the current iteration number.

[0155] In addition, for the individual learning factor c1 and the social learning factor c2 in formula (17), the present invention adopts a similar linear adjustment strategy as follows:

[0156]

[0157]

[0158] Among them, c 1_start and c 2_start They represent the initial values ​​of parameters c1 and c2 before the iteration starts, respectively. 1_end and c2_end Respectively represent the final values ​​of parameters c1 and c2 in the iteration.

[0159] The process of UAVs deployment optimization algorithm based on PSO-GA is as follows:

[0160] Step 1. Randomly initialize the population, including the population size pN and the maximum number of iterations iter max And particle U i , and initialize the local optimal solution pBest of each population i , and the global optimal solution gBest.

[0161] Step 2. When the algorithm execution conditions are met, the particle population is updated through the crossover and mutation operation of the particles, and the calculation offloading scheme is obtained by calling Algorithm 2. The fitness of each particle is calculated according to formula (11), and the local optimal solution pBest is updated according to the fitness. i And the global optimal solution gBest.

[0162] Step 3. Output the final global optimal solution gBest.

[0163] Algorithm 1 PSO-GA algorithm

[0164] / *Find the optimal deployment solution based on PSO-GA algorithm* /

[0165] Input: GTs coordinates u GT , calculate the task size D i , calculation task complexity S i 、Deadline E i

[0166] Output: A global optimal solution gBest that satisfies the conditions

[0167]

[0168]

[0169] 4.2 Inner layer offloading optimization algorithm based on greedy strategy

[0170] The lower layer of the proposed PG-G algorithm needs to achieve computational offloading optimization under the given UAVs deployment scheme, including optimization of user association {α} and offloading rate {β}, as follows:

[0171]

[0172] Since the computing offloading problem in the cloud environment is usually an NP Hard problem, the efficiency of solving the problem using traditional algorithms is low. The Greedy algorithm is a low-complexity approximation algorithm. Its core idea is to always choose the option that currently seems to be the best in the process of solving the problem, so that the approximate optimal solution to the original problem can be quickly obtained. In the model proposed in the present invention, the model constructed according to formula (5) can be obtained: the closer the distance between the GT and the associated UAV, the greater the channel gain between the two, and the greater the corresponding transmission rate. Therefore, associating each GT with the nearest UAV can greatly reduce the transmission delay of the computing task.

[0173] However, since the service range of each UAV is limited, GTs outside the coverage of UAVs cannot offload tasks to UAVs; at the same time, since the parallel processing capability of UAVs is limited, it is impossible to respond to too many requests at the same time. Therefore, the present invention divides all GTs into three types according to the distribution of GTs:

[0174] 1) The GT is not within the coverage of any UAV;

[0175] 2) The GT is within the coverage of only one UAV;

[0176] 3) The GT is located within the coverage area of ​​multiple UAVs.

[0177] Then, we give the priorities of the above three categories in descending order. The first category of GTs has the highest priority, because this type of GTs cannot be associated with UAVs and does not occupy the computing resources of UAVs, so it will not affect the offloading strategy of the other two categories of GTs; the second category of GTs has the second priority, because this type of GTs can only be associated with the nearest UAV; the third category of GTs can select the appropriate UAV to associate with among multiple UAVs, so it has the lowest priority.

[0178] Then, we decide the task offloading rate of GTs. For the first type of GTs mentioned above, since they cannot be associated with UAVs, that is, At this time, the original computing task is only divided into the local execution part and the LEO satellite execution part; for the second and third types of GTs, the original task is divided into the local execution part, the UAV execution part and the LEO satellite execution part. Since the deployment location of UAVs and the relationship between GTs and UAVs are already determined, the task offloading rate optimization problem is a standard linear programming problem and can be directly solved using existing toolkits (such as CVXPY).

[0179] Based on the above analysis, the present invention proposes the following greedy strategy:

[0180] 1) For the first type of GTs, no UAVs are associated;

[0181] 2) Associate the second type of GTs to the nearest UAV. If there is an M max If there are GTs associated with the UAV, the GT with the largest delay among all the GTs associated with the current UAV is selected to make it give up associating with the UAV and adjust the unloading rate;

[0182] 3) For the third type of GTs, always try to associate it with the nearest UAV. max If there are GTs associated with the UAV, the GT with the largest delay among all the third-category GTs associated with the current UAV is selected and unloaded to other nearest UAVs. If it cannot be unloaded to other UAVs, it is made to give up being associated with the UAV and the unloading rate is adjusted.

[0183] The calculation offloading optimization algorithm based on Greedy is as follows:

[0184] Step 1. Divide the GTs into three sets according to the positions between GTs and UAVs.

[0185] Step 2. For the GTs in the first set, the unloading rate is directly calculated. For the GTs in the second and third sets, the association relationship with the UAVs is first determined, and then the unloading rate is calculated.

[0186] Step 3. Output the final relevance set α and uninstall rate set β.

[0187] Algorithm 2 Greedy algorithm

[0188] / *Solve the unloading solution based on the greedy strategy* /

[0189] Input: GTs coordinates u GT , the coordinates u of UAVs UAV , calculate the task size D i , calculation task complexity S i 、Deadline E i

[0190] Output: Relevance set α and uninstall rate set β

[0191]

[0192]

[0193] 2. Advantages and uses of this product.

[0194] In the present invention, a joint optimization algorithm PG-G for UAVs deployment and computing task offloading in an air-space-ground integrated MEC system based on PSO-GA and Greedy algorithm is proposed. The algorithm can converge to a stable value in a finite number of iterations and has a better optimization effect than the benchmark algorithm. The invention is applicable to an air-space-ground integrated MEC system with multiple UAVs and multiple ground devices. By optimizing the deployment position of UAVs and the offloading strategy of computing tasks, the average response time of each computing task in the system is minimized.

[0195] In order to verify the feasibility and effectiveness of the proposed PG-G algorithm, simulation experiments are carried out in this section for the following research questions:

[0196] 1. Does PG-G converge when solving the average response time of the system?

[0197] 2. Is the PSO-GA algorithm on top of PG-G effective?

[0198] 3. Is the Greedy algorithm under PG-G effective?

[0199] 5.1 Experimental Setup

[0200] In this experiment, we consider a 200×200m 2 The CPU computing frequency of GTs distributed in this area is 0.8~1.2GHz, the input data volume of GTs computing tasks follows the random distribution in the range of 10M~15Mbits, the complexity of computing tasks is set to 80~120cycles / bit, and the maximum tolerance delay of tasks is set to 1s. In order to provide network services to GTs, N=4 UAVs are deployed in this area as edge servers, and the CPU computing frequency of UAVs is 2GHz; in addition, a LEO satellite is used as a cloud server, and the computing frequency allocated by the LEO satellite to each GT is 3GHz, and the data transmission rate between the satellite and GTs is set to 20Mbits / s. Other system parameters are shown in Table 1.

[0201] Table 1 System parameter settings

[0202]

[0203]

[0204] 5.2 Convergence of PG-G

[0205] In this section, we show the convergence curves of the proposed PG-G algorithm in different scenarios. Figure 5In the figure, we can see that the proposed PG-G algorithm can always converge to a stable value within 500 iterations. In addition, we can see that when the service range of UAVs is 40m, the convergence speed is relatively slow, while when the service range of UAVs is 80m, the algorithm converges faster. This is because when the service range of UAVs is 40m, there are more GTs that cannot be covered by UAVs. In order to provide services to more GTs, stricter requirements are required for the deployment of UAVs, so it takes more time to search using PG-G; when the service range of UAVs is 80m, almost all GTs can be covered by UAVs, so the time required for PG-G to search is relatively short. In addition, we can see that when the number of users is 20, the convergence speed of PG-G is relatively fast, and when the number of users increases, the convergence speed of the algorithm decreases accordingly.

[0206] 5.3 Effectiveness of the PSO-GA Algorithm on PG-G

[0207] In order to verify the effectiveness of the PSO-GA algorithm on the upper layer of the proposed PG-G algorithm, other algorithms need to be introduced for comparison. However, since the formalized problem is a mixed nonlinear optimization problem, it is impossible to obtain the ideal solution and optimal solution of the problem by conventional means; in addition, since there are few existing studies on air-space-ground integrated MEC systems, and there are many differences between the research of the present invention and the existing work, it is impossible to directly use the methods in the existing work as the comparison algorithm of the present invention. Therefore, this summary introduces the comparison algorithms RAN-G and DE-G by rewriting the two benchmark algorithms commonly used in related research as follows:

[0208] 1) RAN-G: The RAN-G algorithm uses a random algorithm to replace the PSO-GA algorithm on the upper layer of PG-G as the UAVs deployment algorithm, and the lower layer uses the Greedy algorithm proposed in this invention as the computational offloading algorithm. Since the random algorithm has strong uncertainty, in the RAN-G algorithm, the method of repeating 1000 experiments and taking the average is used as the final result of the algorithm.

[0209] 2) DE-G: The DE-G algorithm uses the differential evolution algorithm as the search engine, replacing the PSO-GA algorithm of the present invention as the deployment algorithm of UAVs. The lower layer uses the Greedy algorithm proposed in the present invention as the computational offloading algorithm.

[0210] exist Figure 6 and Figure 7The average response time of each algorithm system when the service radius of UAVs is 40m and 80m is shown in the figure. It can be seen that in various scenarios, the PG-G proposed in the present invention has a significant optimization effect compared with the RAN-G based on the random algorithm, and the effect is significant when the service radius is 80m. This is because when the service radius of UAVs is large, an effective deployment scheme can cover more GTs, so that more GTs can offload part of the tasks to UAVs for execution, thereby making full use of the computing resources of UAVs; and a bad deployment scheme, even if the service range of UAVs is large enough, will make some GTs unable to associate with UAVs, thereby reducing resource utilization. In addition, when the number of terminal devices is 30 and the service radius of UAVs is 80m, DE-G has a better optimization effect. In other scenarios, PG-G has the best optimization effect, so it can be proved that the PSO-GA algorithm is effective in realizing UAVs deployment.

[0211] 5.4 Effectiveness of PG-G Lower Layer Algorithm

[0212] In this section, in order to verify the effectiveness of the Greedy algorithm under the PG-G algorithm, the comparison algorithms PG-RAN and PG-AVG are introduced:

[0213] 1) PG-RAN: PG-RAN adopts the PSO-GA algorithm as the deployment algorithm for UAVs. The lower layer of the algorithm associates with the connectable UAVs in a random manner, and then randomly divides the computing tasks and offloads them to each end for execution.

[0214] 2) PG-AVG: PG-AVG adopts the PSO-GA algorithm proposed in the present invention as the deployment algorithm of UAVs. In the lower layer of the algorithm, it associates with the nearest connectable UAV, and then evenly distributes the computing tasks to each end for execution according to the connection situation.

[0215] Figure 8 and Fig. 9 The average response time of each algorithm system when the service radius of UAVs is 40m and 80m is shown respectively. As can be seen from the figure, the proposed PG-G algorithm has the best performance in various scenarios, followed by PG-AVG, and PG-RAN is the worst. This is because compared with the random calculation-based offloading strategy, the Greedy algorithm proposed in the present invention has been optimized in both user association and offloading rate decision-making, so the performance is significantly improved; and compared with PG-AVG, the Greedy algorithm proposed in the present invention is significantly better than the average distribution scheme in offloading rate optimization. Therefore, when the service range of UAVs is large, the Greedy algorithm proposed in the present invention can achieve better optimization effects.

Claims

1. The deployment optimization and computation offloading method of the integrated air-ground-space mobile edge computing system is characterized in that The proposed system consists of a ground layer of M GTs, an aerial layer of N UAVs, and a space layer consisting of a LEO satellite. The sets of GTs and UAVs are represented as M = {1, 2, ..., M} and N = {1, 2, ..., N}, respectively. In the ground layer, GTs with limited computing power are distributed in areas lacking communication infrastructure and perform tasks with computing requirements. UAVs in the aerial layer serve as edge nodes and provide edge computing services for GTs. In the space layer, LEO satellites provide centralized cloud computing services for all GTs in their coverage area. GTs are allowed to offload part of their computing tasks to UAVs and LEO satellites for execution. The three-dimensional Euclidean coordinate system is used to represent the positions of ground terminal devices and UAVs. The position of the i-th GT in the current time slot is represented by the three-dimensional coordinate system Indicates that, accordingly, the current position of the jth UAV is expressed in three-dimensional coordinates Where H represents the flight altitude of UAVs; Take a 4-tuple I i = <D i ,S i ,O i ,E i > represents the computing task that the i-th GT needs to perform in the current time slot, where D i Represents computing task I i The input data size; S i Indicates the number of computing resources required to execute the unit computing task input; i Indicates the size of the task execution result; E i Indicates the maximum tolerable delay of the computing task; the computing task execution result O i The size of D is much smaller than the input of the task. i , computing task I i The completion delay shall not exceed the maximum tolerable delay E i ; The calculation task I i The partial offloading strategy is adopted to allow GTs to divide their computing tasks and offload part of the computing tasks to UAVs and LEO satellites for execution. The proportion of the local execution part and the subtasks offloaded to UAVs and LEO satellites are respectively expressed as and To express; When the i-th GT is not within the coverage of any UAV, the GT cannot offload part of the computing tasks to the UAVs for execution. According to the different execution locations of each subtask, the execution process of the original computing task is divided into three different modes, namely: local computing mode executed locally, edge computing mode executed on UAV, and cloud computing mode executed on LEO satellite; these three computing modes are executed in parallel; 1) In local computing mode, define parameters is the computing frequency of the local CPU of the i-th GT, according to the computing task I i And local uninstall ratio The definition of the local task size to be processed is Therefore, the delay caused by local calculation is expressed as 2) In edge computing mode, the latency of executing computing tasks will consist of three parts, namely: the transmission latency of offloading the task to the associated UAV, the computing latency of executing on the UAV, and the backhaul latency of returning the computing results to the GT; The wireless channel between the i-th GT and the j-th UAV is modeled as a line-of-sight channel link model; the wireless channel power gain between the i-th GT and the j-th UAV is quantified as Where h0 represents the channel gain in the wireless channel when the reference distance d0 = 1 meter, d ij It represents the distance between the i-th GT and the j-th UAV, which can be expressed as Therefore, the data transmission rate between the i-th GT and the j-th UAV is further expressed as Where B represents the system bandwidth between GTs and UAVs, P i represents the transmission power of the i-th GT; σ 2 Represents the background noise power in the channel; Define a binary variable β ij represents the relationship between GTs and UAVs, where β ij =1 means that the i-th GT decides to offload part of the task to the j-th UAV for execution, otherwise β ij = 0; it is stipulated that each GT can at most offload part of the task to one UAV for execution, so there are the following constraints: In addition, the upper limit of the number of concurrent operations per UAV in the current time slot is specified, so there are the following constraints: For the i-th GT, its task completion delay in edge computing mode consists of two parts: transmission delay and calculation delay, which can be specifically expressed as in, It represents the CPU computing frequency assigned to the associated GT by the jth UAV; 3) In the cloud computing mode, GTs and LEO satellites will use direct ground-to-space transmission technology to offload some computing tasks to remote execution, and the data transmission rate between all GTs and LEO satellites is uniformly defined as R s , define the parameters represents the CPU computing frequency allocated to each GT by the LEO satellite. Therefore, the task completion delay in the cloud computing mode is expressed as Since the local computing mode, edge computing mode, and cloud computing mode can be executed in parallel, for the i-th GT, the completion delay of its computing task is T i Should be expressed as The system average response delay is defined as the average of the completion delays of all computing tasks in the system, which is expressed as follows: The optimization problem of joint deployment and computation offloading is formulated as: Among them, C1 means that the completion delay of each task cannot exceed the maximum tolerable delay of the task; C2 and C3 represent the constraints of the task offloading rate; C4-C6 represent the association constraints between GTs and UAVs; C7 means that each UAV can only provide offloading services for GTs within a limited range; The following encoding mechanism is adopted: the population is composed of several particles, each particle represents a deployment plan of a UAVs cluster; each particle is encoded as a set of N two-dimensional vectors, each of which represents the horizontal coordinates of a UAV in the current time slot; specifically, in the tth iteration, the kth particle in the population is represented as in It represents the horizontal coordinate of the jth UAV in the t-round iteration, which is specifically expressed as in and They represent the horizontal and vertical coordinates of the deployment position of the jth UAV in the kth particle, respectively; The PSO-GA algorithm improves the particle update process of the traditional PSO algorithm by introducing the crossover and mutation operators in the GA algorithm. At this time, the update method of the kth particle is changed to Among them, M u () represents the mutation operator of the PSO-GA algorithm, C p () and C g () represents the crossover operator; The inertia part of the particle is updated as follows: Among them, r1 is a random value in the interval [0,1]; Indicates randomly selected particles A certain quantile on , and perform random mutation within the threshold range; In the PSO-GA algorithm, the traditional individual learning part is combined with the social learning part and the crossover operator of the GA algorithm. The corresponding update methods are: Among them, r2 and r3 are two random values ​​in the interval [0,1]; the crossover operator randomly selects two quantiles on the particle and uses pBest k Or replace the values ​​between the corresponding quantiles on gBest; A linear inertia weight factor adjustment strategy is adopted, which will make the inertia weight factor w decrease linearly with the increase of the number of iterations. The specific adjustment strategy is as follows: In the formula, w max and w min They are the maximum and minimum values ​​set by w, iters max Indicates the agreed maximum number of iterations, iters cur Indicates the current iteration number; A similar linear adjustment strategy is used as follows: Among them, c 1_start and c 2_start They represent the initial values ​​of parameters c1 and c2 before the iteration starts, respectively. 1_end and c 2_end Respectively represent the final values ​​of parameters c1 and c2 in the iteration; The process of UAVs deployment optimization algorithm based on PSO-GA is as follows: Step 1: Randomly initialize the population, including the population size pN and the maximum number of iterations iter max And particle U i , and initialize the local optimal solution pBest of each population i , and the global optimal solution gBest; Step 2: When the algorithm execution conditions are met, the particle population is updated through the crossover and mutation operations of the particles, and Algorithm 2 is called to obtain the computational offloading scheme. The fitness of each particle is calculated according to formula (11), and the local optimal solution pBest is updated according to the fitness. i And the global optimal solution gBest; Step 3: Output the final global optimal solution gBest; The proposed PSO-GA algorithm lower layer needs to achieve computational offloading optimization under the given UAVs deployment scheme, including user association { α } and uninstall rate {β} optimization, as follows: All GTs are divided into three types according to their distribution: 1) The GT is not within the coverage of any UAV; 2) The GT is within the coverage of only one UAV; 3) The GT is within the coverage of multiple UAVs; Then, we give the priorities of the above three categories in descending order; the first category of GTs has the highest priority, because this type of GTs cannot be associated with UAVs and does not occupy the computing resources of UAVs, so it will not affect the offloading strategy of the other two categories of GTs; the second category of GTs has the second priority, because this type of GTs can only be associated with the nearest UAV; the third category of GTs selects the appropriate UAV to associate with among multiple UAVs, so it has the lowest priority; Then, we decide the task offloading rate of GTs; for the first type of GTs mentioned above, since they cannot be associated with UAVs, that is, At this time, the original computing task is only divided into the local execution part and the LEO satellite execution part; for the second and third types of GTs, the original task is divided into the local execution part, the UAV execution part and the LEO satellite execution part.

2. The method for deployment optimization and computation offloading of an air-ground integrated mobile edge computing system according to claim 1, characterized in that: The calculation offloading optimization algorithm based on Greedy is as follows: 1) Divide the GTs into three sets according to the position between GTs and UAVs; 2) For the GTs in the first set, the unloading rate is directly calculated. For the GTs in the second and third sets, the association relationship with the UAVs is first determined, and then the unloading rate is calculated; 3) Output the final association set α and the set of uninstall rates β.

Images