A computational offloading method for distributed deep learning in a satellite-ground collaborative network

Abstract

The present invention belongs to the field of communication technology, and discloses a computational offloading method for distributed deep learning in a satellite-ground collaborative network, including: designing a cloud-edge-local three-layer architecture for an on-orbit edge computing scenario, where ground users' tasks can be executed locally, on a low-orbit satellite edge server, or on a ground cloud computing center; modeling offloading decisions and resource allocation problems as mixed integer programming problems; constructing a distributed offloading algorithm for a satellite-ground collaborative network to solve the mixed integer programming problem in steps and obtain the optimal offloading decision and resource allocation results. The goal of the problem described in the present invention is to minimize the weighted total cost of latency energy consumption in the architecture while satisfying constraints. To solve the above problems, the present invention proposes a CNDO algorithm, which obtains the optimal offloading decision in the TLOE scenario in two steps, avoiding the problem of the curse of dimensionality. Experimental results show that the scheme described in the present invention has good performance.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a computational offloading method for distributed deep learning in a satellite-ground collaborative network. background

[0002] With the deepening of 6G technology research, the traditional communication industry structure is undergoing a gradual and profound change. The popularity of 5G networks and IoT devices has greatly increased the demand of global users for emerging applications such as real-time transmission of high-definition video, virtual reality, and autonomous driving. Communication networks are gradually developing towards the vision of full coverage, rapid response, and anytime, anywhere user access. A large number of resource requests are pouring into the network, posing new challenges to the limited distribution of computing resources in traditional communications.

[0003] Due to the geographical restrictions of complex terrains such as mountains and deserts and remote areas such as rural areas, traditional communication equipment has been unable to meet the QoS requirements of smart devices in various places. Under the influence of natural disasters, ground communication networks are easily damaged and cause large-scale communication interruptions. At the same time, the distance between existing cloud computing centers and mobile terminals is far away, making it difficult to meet the latency requirements of emerging applications.

[0004] The emergence of some emerging technologies has gradually broken the above dilemma. The emergence of Mobile Edge Computing (MEC) networks has directly sunk computing cache resources to the edge of users, breaking the distance limit between cloud servers and users. A large satellite constellation composed of several low-orbit satellites (Low Earth Orbit, LEO) can meet the requirements of global regional coverage. It has the characteristics of low communication cost, miniaturized equipment, and low transmission delay. The propagation delay can reach milliseconds. The satellite-to-ground communication system derived from low-orbit satellites combines the advantages of ground communication facilities and satellite networks, which can supplement and expand the ground network and make up for the limitations of existing facilities.

[0005] Satellite communications are currently developing in the direction of broadband networks, dense coverage, intelligent applications, and space-ground linkage. Similar to the computing offloading optimization strategy in vehicle-mounted edge computing or drone assistance. For example, "Y.Wang, J.Zhang, X.Zhang, P.Wang and L.Liu, "A Computation Offloading Strategy in SatelliteTerrestrial Networks with Double Edge Computing," 2018IEEE InternationalConference on Communication Systems (ICCS), pp.450-455, 2018." and "J.Zhang, X.Zhang, P.Wang, L.Liu and Y.Wang, "Double-edge intelligent integrated satelliteterrestrial networks," in China Communications, vol.17, no.9, pp.128-146, Sept.2020." provide a model of a dual-edge satellite-ground fusion network, deploying MEC servers on ground base stations and LEO satellites to ensure various needs including computing-intensive and delay-sensitive services. In “Y.Wang, J.Zhang, X.Zhang, P.Wang and L.Liu, “A Computation Offloading Strategy in Satellite Terrestrial Networks with Double Edge Computing,” 2018IEEE International Conference on Communication Systems (ICCS), pp.450-455, 2018.”, the Hungarian algorithm is used to solve the problem of computation offloading. “J.Zhang, X.Zhang, P.Wang, L.Liu and Y.Wang, “Double-edge intelligentintegrated satellite terrestrial networks,” in China Communications, vol.17, no.9, pp.128-146, Sept.2020.”, the caching and delivery problems of satellite-to-ground network terminals are mainly considered, and DQN is used for modeling to reduce the average service delay of the system.“Y.Wang, J.Yang, X.Guo and Z.Qu, “A Game-TheoreticApproach to Computation Offloading in Satellite Edge Computing,” in IEEE Access, vol.8, pp.12510-12520, 2020.” proposed a game theory-based computing offloading strategy for existing satellite-to-ground networks. The above literature considers the local-satellite two-layer model, while ignoring the abundant computing resources of ground cloud computing centers. “Q.Tang, Z.Fei, B.Li and Z.Han, “Computation Offloading in LEO Satellite Networks With Hybrid Cloud and Edge Computing,” in IEEE Internet ofThingsJournal, vol.8, no.11, pp.9164-9176, Jul.2021.” considers the problem of satellite coverage limitations. It constructs a three-layer model that takes into account the situation with satellite assistance and provides a solution to minimize energy consumption using the ADMM algorithm.

[0006] While on-orbit edge computing provides convenient computing solutions for end users, it also has some difficulties. First, the limited channel capacity of the link will lead to a non-negligible transmission delay between the satellite and the ground. Therefore, the problem of finding a reasonable bandwidth allocation and the best offloading decision needs to be solved urgently. Secondly, the designed offloading decision and resource allocation problem is a mixed integer programming (Mixed-Integer Programming, MIP) problem. Traditional methods such as dynamic programming and branch-and-bound algorithms are not suitable for MEC networks with a large number of users. Using traditional methods such as game theory and Hungarian algorithm to obtain the optimal result requires a lot of iterations, which is difficult to meet the low latency requirements of ground users such as vehicles. The potential and feasibility of deep learning in terrestrial edge computing provide new directions for research.

[0007] With the rapid development of the Internet of Things (IoT) and its derivative technologies such as the Aerial Internet of Things (AIoT), the number of smart devices and their computing needs have exploded. This makes it difficult for existing ground base stations to guarantee users' computing needs and low latency requirements. The edge computing architecture of on-orbit satellites provides a feasible solution to the above challenges. For the existing on-orbit edge computing system, there are still the following problems to be solved: when ground user equipment performs task offloading, if the task size is not considered and all tasks are offloaded to edge devices, it will lead to congestion of edge CPU requests and waste of local computing resources. At the same time, the rich computing resources of the ground cloud computing center are also ignored. Moreover, the task offloading of ground users is a dynamic process. At present, most traditional methods solve the optimization problem at a certain moment. They do not consider the research on the dynamic computing offloading process of ground users in continuous time, and it is difficult to connect with practical applications.

[0008] Through the above analysis, the problems and defects of the prior art are as follows:

[0009] (1) The limited channel capacity of the link will result in a non-negligible transmission delay between the satellite and the ground.

[0010] (2) The designed offloading decision and resource allocation problem is a MIP problem. Traditional methods such as dynamic programming and branch-and-bound algorithms are not applicable to MEC networks with a large number of users.

[0011] (3) Using traditional methods such as game theory and the Hungarian algorithm to obtain the optimal result requires a large number of iterations, which is difficult to meet the low latency requirements of ground users such as vehicles.

[0012] (4) With the explosive growth in the number of smart devices and their computing needs, existing ground base stations are unable to meet users’ computing needs and low latency requirements.

[0013] (5) When ground user equipment performs task offloading, if the task size is not considered and all tasks are offloaded to edge devices, it will lead to congestion of edge CPU requests and waste of local computing resources. At the same time, the rich computing resources of the ground cloud computing center are also ignored.

[0014] (6) Task offloading for ground users is a dynamic process. Currently, most traditional methods solve optimization problems at a certain moment. They do not consider the dynamic computation offloading process of ground users in continuous time, which makes it difficult to integrate with practical applications.

[0015] The difficulty of solving the above problems and defects is as follows: each of the above problems has a certain degree of processing difficulty. Regarding problem (1), low latency is an important requirement for future communication systems, and solving the problem of excessive latency has always been a difficult problem for researchers when designing communication systems. In view of the increasing number of emerging applications, proposing new solutions is an important difficulty and major challenge in adapting to application innovation. (5) and (6) are the bottlenecks of the current traditional solutions. To ensure that existing ground computing resources are not wasted and task offloading conforms to the dynamic process in actual applications, it is a difficult problem that we urgently need to overcome in processing satellite-ground systems. In summary, the above problems have a certain degree of processing difficulty for researchers in this field. The present invention intends to provide corresponding solutions to the above problems. The significance of solving the above problems and defects is that the above problems and defects are important breakthroughs in building a satellite-assisted on-orbit edge computing network and promoting it to enter the practical application stage. Today, when 5G networks provide people with a wide range of communication services, users' communication needs and application types are gradually increasing. Traditional ground networks are no longer sufficient to make up for the supply and demand gap caused by huge application needs. Low-orbit satellite-assisted satellite-to-ground communication technology is listed by industry insiders as an important entry point and main enabling technology for entering the 6G era due to its wide coverage of the ground. At present, the biggest difference between 6G communication technology and 5G technology is that 6G, based on 5G, has drawn a grand blueprint for seamless global coverage. This goal is difficult to achieve only with existing ground facilities. Therefore, the successful deployment of non-ground equipment-assisted communication networks is of great significance for human society to enter the 6G era. Summary of the invention

[0016] In view of the problems existing in the prior art, the present invention provides a distributed deep learning computing offloading method in a satellite-ground collaborative network.

[0017] The present invention is implemented as follows: a method for offloading calculations of distributed deep learning in a satellite-ground collaborative network, the method comprising the following steps:

[0018] Step 1: System model construction: Design a cloud-edge-local three-layer architecture network (Three Layers Orbital Edge computing architecture, TLOE) for on-orbit edge mission computing to build a system model. The task can be selectively executed directly locally or offloaded to the LEO satellite edge (Satellite Edge Server, SatEC) server or ground cloud computing center for execution;

[0019] Step 2: Problem formulation: Model the offloading decision and resource allocation problem as a mixed integer programming problem according to the task focus;

[0020] Step 3: Compute offloading: Construct a collaborative satellite-terrestrial network distributed offloading algorithm (CNDO) to obtain the optimal offloading decision and resource allocation results.

[0021] Step 1 to Step 3 constitute this solution and are important links in the implementation of this solution. The coordinated implementation of the three steps provides a new solution for the existing satellite-ground collaborative network and provides a solution combining deep learning for the development of on-orbit edge computing.

[0022] Furthermore, in step 1, the system model is constructed, including:

[0023] The constructed three-layer on-orbit edge computing network TLOE includes a ground central cloud server, L LEO satellites and N ground users (Terrestrial Users, TU), denoted as and Each LEO satellite is equipped with a MEC server to accept tasks from ground users and complete calculations. At the same time, the LEO satellite can also serve as a relay device for the cloud server to forward the task data of ground users. Each ground user has a task to be calculated in the current period. When the current task of a user is calculated, the next task will be generated. All tasks are inseparable and can only be unloaded as a whole.

[0024] The details of each task are grouped by variables {D i in , δ i} description, the two variables in the variable group represent the task input [in bits] and task processing density [in cycles / bit] of the i-th TU. Since the data volume of most task calculation results is small, the model does not consider the return process of the calculation results. Therefore, the total number of CPU cycles required to complete the task calculation of the i-th TU is described as Z i =D i in δ i , where D i in and δ i The value of is obtained by analyzing the execution of each task. The computational density of different tasks may not be equal. To simplify the calculation, the same computational density can be used for the tasks of each TU.

[0025] In the TLOE framework, all available CPU resources are defined as the set of each task j∈{1, 2, 3, ..., L, L+1, L+2}, where {1, 2, 3, ..., L} represents all available LEO satellite edge servers, L+1 represents the cloud server located on the ground, and L+2 represents the task directly computed on the local CPU. The binary computation offloading decision of the task of the i-th TU calculated by the j-th CPU is denoted as x ij If the ith user task is executed on the corresponding satellite edge server numbered j, then x ij is 1, otherwise it is 0. i，L+1 1 means that the ith task is executed by the cloud server, x i，L+2 When it is 1, it indicates that the task is calculated locally.

[0026] Assuming that each ground user can only access one LEO satellite for data transmission, multiple ground users share the same spectrum resources, which means that there is mutual interference between ground users. Therefore, the uplink transmission rate from the i-th TU to the j-th satellite edge server is calculated as:

[0027]

[0028] Wherein, the channel power gain between the i-th TU and the j-th satellite edge server is described as h ij ; The transmit power of the i-th TU is expressed as σ 2 represents the noise power, B ij Represents the bandwidth allocated by the j-th LEO satellite to the i-th TU.

[0029] Furthermore, in step 1, the system model construction further includes:

[0030] (1) Local computing model: Due to the limited computing resources and power of mobile devices, the device allocates basic tasks to perform calculations locally as much as possible. The time spent on the local CPU processing of the task of the i-th TU is expressed as:

[0031]

[0032] in, is the local CPU clock frequency of the i-th TU; the execution power of the i-th TU in local task calculation is expressed as:

[0033]

[0034] Among them, k i represents the effective switch capacitance; V is a positive constant, set to 3; the corresponding energy consumption when the i-th TU adopts the local calculation method is:

[0035]

[0036] (2) Satellite edge computing model: When using satellite edge computing methods for offloading calculations, the total processing delay should include three parts, namely, the transmission delay of the transmission task data Propagation delay at the speed of light Computational latency of tasks on satellite edge servers Although the satellite is a low-orbit satellite, the distance between it and the TU is much greater than the distance between TUs on the ground. In this TLOE model, the inclination between the satellite and the ground TU is not considered. It is assumed that the actual distance between the TU and the satellite is consistent with the altitude of the LEO satellite above the ground, which is set to a constant H. If the i-th TU needs to offload the task to the satellite edge server J, the total processing time is:

[0037]

[0038] where C is the speed of light, and is the CPU clock rate of the jth satellite edge server used to calculate the i-th TU task. The energy consumption of the low-orbit satellite offloading calculation process is described as:

[0039]

[0040] (3) Cloud computing model: The central cloud server is located on the earth, and its computing power is much greater than that of the satellite edge server and the local CPU. Before the cloud server provides computing services to users, it needs to forward user tasks to the cloud server via satellite as a relay.

[0041] Similar to the delay model and energy consumption model of the LEO satellite edge server, the task calculation of the cloud server only has one more process of forwarding from the satellite to the cloud than the task calculation of the satellite edge server. The forwarding delay is expressed as In addition to ignoring the downlink propagation delay, the return process of the download task is also ignored, so the delay and energy consumption model is expressed as:

[0042]

[0043] The energy consumption of the cloud computing model is:

[0044]

[0045] Among them, R EC is the uplink transmission rate from edge server to cloud server, f C represents the CPU computing rate of the cloud server, and the transmission power of the satellite edge server is P EC Therefore, for the task of the i-th TU, the total delay of task calculation is expressed as:

[0046]

[0047] The total energy consumed by the task is obtained from the following formula:

[0048]

[0049] Further, in step 2, the problem is formulated, including:

[0050] According to the system model, for the TLOE calculation system, the total cost of task processing is expressed as the weighted sum of energy consumption and delay when processing all tasks, which is expressed as the following cost function:

[0051]

[0052] Among them, the function The variables satisfy

[0053] Weight parameter satisfy The cost function needs to be adjusted according to the specific requirements of TU. For latency-sensitive tasks, If users pay more attention to reducing the total energy consumption of the system, In extreme cases, when When , only the minimum total delay of TU is considered. When , the delay of TU is ignored and only the optimal energy consumption is considered.

[0054] Formulate the optimization problem and use the proposed CNDO algorithm to solve it. There are the following optimization problems:

[0055]

[0056]

[0057]

[0058]

[0059] x ij ∈{0, 1}

[0060] Among them, the first two constraints are bandwidth limitations, Indicates that the allocated bandwidth must be a non-negative number. Indicates that the total bandwidth allocated to all users must be less than the maximum bandwidth of the satellite edge server. It means that for a task of a TU, only one CPU can be used, including the local CPU of the TU; the offload variable is a binary integer variable. The optimization problem is a MIP problem in a high-dimensional state, which is non-convex and NP-hard. It is difficult to solve it using a traditional heuristic search algorithm. Therefore, a distributed deep learning algorithm is introduced to solve this problem in two stages.

[0061] Furthermore, in step 2, the task focuses on:

[0062] According to the system model, for the cloud-edge-local three-layer architecture network proposed in the present invention, the execution of tasks should be carried out with emphasis for different task types or different application scenarios; wherein the emphasis includes the following two types: delay-sensitive and energy-sensitive;

[0063] The delay-sensitive type focuses more on reducing the total delay of the system's computing offloading process; the energy-sensitive type focuses more on having lower total energy consumption in the computing offloading process; the task focus can be adjusted by setting weight parameters.

[0064] Further, in step 3, the construction of the distributed offloading algorithm CNDO for the collaborative satellite-ground network includes:

[0065] In the CNDO architecture, S suboptimal decisions are obtained in the first stage. In the second stage, these suboptimal decisions are used as known quantities to obtain the optimal decision by solving the bandwidth allocation problem in the original optimization problem and completing the update of the data pool data at the same time; the CPU workload is As the input of CNDO, it outputs the optimal offloading decision x after the network runs * ; The results of each operation are combined and stored in the data pool. When the data pool is full, the old data is continuously overwritten with the updated data to train the DNN to solve the MIP problem.

[0066] Further, in step 3, the construction of the distributed offloading algorithm CNDO for the collaborative satellite-ground network also includes:

[0067] (1) Generation of uninstall decision options

[0068] In the model, the original unloading decision space is too large, that is, x * ∈{0, 1} N(L+2) ； The goal of this invention is to obtain a suitable offloading strategy function π, which generates the current optimal offloading action for each DNN module in the entire distributed structure. The current optimal offloading actions of all DNN modules constitute S suboptimal offloading actions for the entire network. As a reward to the network, the input task computation amount is the state. The calculation process is performed once in each time slot t, and there are: t∈{1, 2, ..., T}. The decision obtained by each DNN network is the optimal offloading action under this DNN. Using S DNN modules with the same structure, Although the S DNN modules have the same structure, the network parameters updated during the training process are different. different.

[0069] (2) Resource allocation and data pool update

[0070] The second phase of the algorithm includes resource allocation, optimal action generation, and data pool update. Resource allocation is to allocate available bandwidth to different users so that the weighted cost Minimum.

[0071] When these candidate binary offloading decisions are obtained through the calculation in the first stage, it is considered that the offloading decision space has been expanded from 2 N(L+2) Reduce to S; you need to choose one of these S actions that satisfies the minimum unloading cost As the final optimal action x * Each suboptimal unloading decision is considered as a known quantity in the original problem. Now we only need to solve the resource allocation result The problem now is a convex optimization problem. The CVXPY tool is used to calculate resource allocation As a result, the weighted cost Obtained by the formula.

[0072] The network input and optimal offloading decision will be concatenated as and stored in the data pool. The data pool is initially empty and has a size of μ. The network randomly generates parameters to complete initialization. As the time slot t increases, the network will continue to fill When the data pool is full, the oldest data will be discarded and replaced by new data. All DNN blocks can randomly select several items in the data pool to train the network. Here, cross entropy is used as the loss function, and other loss functions can also be selected according to the specific model;

[0073] The CNDO algorithm process of the cloud-edge-local three-layer architecture is as follows:

[0074] 1) Input: CPU cycles of all TU tasks in each time frame

[0075] 2) Output: The best offloading decision for each frame

[0076] 3) Initialization:

[0077] ① Initialize all DNN parameters with random parameters

[0078] ② Clear the data pool of size μ;

[0079] ③For each time frame t, t=1, 2, ..., T, do the following operations

[0080] ④ Input to SDNNs;

[0081] ⑤ Generate candidate offloading decisions from the Sth DNN

[0082] ⑥Solve the optimization problem and use Get resource allocation plan Compare the results and select the best one

[0083] ⑦ If there is available space in the data pool, store the combined data to the data pool;

[0084] ⑧Otherwise, overwrite the old data with the new data

[0085] ⑨ Randomly extract S batches of data from the data pool to train DNN, and update network parameters during training

[0086] Another object of the present invention is to provide a distributed deep learning computing unloading system in a satellite-ground collaborative network using the distributed deep learning computing unloading method in a satellite-ground collaborative network, wherein the distributed deep learning computing unloading system in a satellite-ground collaborative network comprises:

[0087] System model building module for designing cloud-edge-local three-layer architecture network TLOE for on-orbit edge mission computing, which is executed locally, on SatEC servers or in ground cloud computing centers;

[0088] Problem formulation module, which is used to model the offloading decision and resource allocation problem as a MIP problem;

[0089] The computational offloading module is used to obtain the optimal offloading decision and resource allocation results by constructing a distributed offloading algorithm CNDO for the collaborative satellite-ground network.

[0090] Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0091] (1) System model construction: Design a cloud-edge-local three-layer architecture network TLOE for on-orbit edge mission computing, which is executed locally, on SatEC servers, or in ground cloud computing centers;

[0092] (2) Problem formulation: The offloading decision and resource allocation problem is modeled as a MIP problem;

[0093] (3) Computational offloading: Construct a distributed offloading algorithm CNDO for collaborative satellite-ground networks to obtain the optimal offloading decision and resource allocation results.

[0094] Another object of the present invention is to provide an information data processing terminal, which is used to implement the computing offloading system of distributed deep learning in the satellite-ground collaborative network.

[0095] Combining all the above technical solutions, the advantages and positive effects of the present invention are as follows: Based on the ground communication network, the present invention conforms to the development vision of the 6G network and proposes a satellite-ground three-layer hybrid architecture. This architecture breaks through the limitations of the existing satellite-ground network that ignores the rich computing resources of the cloud computing center or only uses satellites as auxiliary relay devices, overcomes the application bottleneck of satellite communications in the industry, and creatively proposes to use satellite edge computing as a link in the three-layer satellite-ground network, providing new research ideas for the deployment of satellite communication systems in existing communication systems.

[0096] The computational offloading method of distributed deep learning in the satellite-ground collaborative network provided by the present invention designs a computational offloading strategy for users to minimize the weighted sum of latency and energy consumption in the cloud-edge-local three-layer hybrid AIoT architecture. Tasks can be calculated locally, on the satellite side, or forwarded to the cloud computing center; a satellite-ground network collaborative distributed offloading algorithm (CNDO) based on parallel neural networks is proposed, and multiple optimal offloading decisions in AIoT are given. Experimental results show that this scheme has better performance than other comparison schemes.

[0097] The present invention models the offloading decision and resource allocation problem as a MIP problem and proposes a distributed offloading algorithm (CNDO) for collaborative satellite-ground networks to solve this problem and obtain computational offloading decisions. The proposed model is simulated and compared with other baseline solutions to demonstrate that the proposed solution has good performance in terms of computational performance.

[0098] The present invention proposes a CNDO algorithm based on a three-layer cloud-edge-local satellite-assisted edge computing architecture to minimize the weighted cost of system energy consumption and latency. Using a distributed network composed of multiple parallel DNNs to make unloading decisions can effectively reduce the search space of unloading strategies and avoid the problem of dimensionality curse. For the difficult MLP problem, a limited number of unloading decision variables are found through a deep learning algorithm, and the final calculation results are obtained in two stages. Simulation experiments show that the proposed strategy can produce better unloading decisions, verifying the good performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0100] Figure 1 It is a flow chart of a computational offloading method for distributed deep learning in a satellite-ground collaborative network provided by an embodiment of the present invention.

[0101] Figure 2 It is a structural block diagram of a computing offloading system for distributed deep learning in a satellite-ground collaborative network provided by an embodiment of the present invention;

[0102] In the figure: 1. System model building module; 2. Problem formulation module; 3. Computation offloading module.

[0103] Figure 3 It is a schematic diagram of a collaborative task offloading framework for computing offloading provided by an embodiment of the present invention.

[0104] Figure 4 It is a schematic diagram of the satellite-ground collaborative distributed offloading (CNDO) algorithm architecture provided by an embodiment of the present invention.

[0105] Figure 5 It is a convergence curve diagram when 6 DNNs are used under the distributed offloading algorithm provided in an embodiment of the present invention.

[0106] Figure 6 It is a schematic diagram of normalized performance comparison of different methods provided in the embodiments of the present invention. DETAILED DESCRIPTION

[0107] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0108] In view of the problems existing in the prior art, the present invention provides a computational offloading method for distributed deep learning in a satellite-ground collaborative network. The present invention is described in detail below in conjunction with the accompanying drawings.

[0109] like Figure 1 As shown, the method for offloading computing of distributed deep learning in a satellite-ground collaborative network provided by an embodiment of the present invention includes the following steps:

[0110] S101, System Model Construction: Design a cloud-edge-local three-layer architecture network TLOE for on-orbit edge mission computing, which is executed locally, on SatEC servers, or in ground cloud computing centers;

[0111] S102, Problem Formulation: Model the offloading decision and resource allocation problem as a MIP problem;

[0112] S103, computational offloading: construct a distributed offloading algorithm CNDO for the collaborative satellite-ground network to obtain the optimal offloading decision and resource allocation results.

[0113] like Figure 2 As shown, the computing offloading system for distributed deep learning in a satellite-ground collaborative network provided by an embodiment of the present invention includes:

[0114] System model building module 1 is used to design a cloud-edge-local three-layer architecture network TLOE for on-orbit edge mission computing, which is executed locally, on SatEC servers, or in ground cloud computing centers;

[0115] Problem formulation module 2, used to model the offloading decision and resource allocation problem as a MIP problem;

[0116] The calculation offloading module 3 is used to obtain the optimal offloading decision and resource allocation result by constructing a distributed offloading algorithm CNDO for the collaborative satellite-ground network.

[0117] The technical solution of the present invention is further described below in conjunction with specific embodiments.

[0118] 1. With the rapid development of the Internet of Things (IoT) and its derivative technologies such as the Aerial Internet of Things (AIoT), the number of smart devices and their computing needs have exploded. This makes it difficult for existing ground base stations to guarantee users' computing needs and low latency requirements. The edge computing architecture of in-orbit satellites provides a feasible solution to the above challenges. The present invention designs a computing offloading strategy for users to minimize the weighted sum of latency and energy consumption in the cloud-edge-local three-layer hybrid AIoT architecture. Tasks can be calculated locally, on the satellite side, or forwarded to a cloud computing center. Then, a satellite-ground network collaborative distributed offloading algorithm (CNDO) based on parallel neural networks is proposed, which gives multiple optimal offloading decisions in AIoT. Experimental results show that this scheme has better performance than other comparison schemes.

[0119] The present invention designs a three-layer architecture network for on-orbit edge mission computing. Figure 3 As shown in the figure, users can choose to execute tasks locally or forward them to satellite edge servers or cloud servers to complete the calculation of tasks remotely. In order to obtain the optimal offloading decision and resource allocation results, this paper adopts a computing offloading algorithm based on distributed deep learning. The main contributions of this study are summarized as follows:

[0120] The present invention considers a collaborative cloud-edge-local three-layer on-orbit edge computing architecture (TLOE), where tasks are indivisible and can be performed locally, on SatEC servers, or on ground cloud computing centers. The goal of the present invention is to minimize the weighted total cost of energy consumption and latency of the entire system while jointly optimizing offloading decisions and resource allocation.

[0121] The present invention models the offloading decision and resource allocation problem as a MIP problem, and proposes a distributed offloading algorithm (CNDO) for a collaborative satellite-ground network to solve this problem and obtain a computational offloading decision.

[0122] The proposed model is simulated and compared with other baseline solutions to demonstrate the superior performance of the proposed scheme in terms of computational performance.

[0123] The rest of the paper is organized as follows. In the second subsection, the paper introduces the system model and formulates the optimization problem. The proposed algorithm based on distributed deep learning to generate optimal binary offloading decisions is introduced in the third subsection. The simulation results are given in the fourth subsection.

[0124] 2. System model and problem formulation

[0125] The present invention proposes a collaborative task offloading framework. The device can choose to perform task calculations locally or completely offload the tasks to a cloud server or LEO satellite edge server for calculation.

[0126] 2.1 System Model

[0127] The three-layer on-orbit edge computing network TLOE described in the present invention is as follows Figure 3 As shown in Figure 1, it includes a ground central cloud server, L LEO satellites and N ground users (TU), which can be recorded as and Each LEO satellite is equipped with a MEC server, which can accept tasks from ground users and complete calculations. At the same time, LEO satellites can also serve as relay devices for cloud servers to forward ground user task data. Each ground user has a task to be calculated in the current period. The model considers that when the current task of a user is calculated, the next task will be generated. All tasks are indivisible and can only be unloaded as a whole.

[0128] The specific details of each task can be used with the variable group {D i in , δ i}, the two variables in the variable group are the task input data size of the i-th TU [in bits] and the task processing density [in cycles / bit]. Since the data volume of most task calculation results is small, the model does not consider the return process of the calculation results. Therefore, the total number of CPU cycles required to complete the task calculation of the i-th TU can be described as z i =D i in δ i These D i in and δ i The value of can be obtained by analyzing the execution of each task. The computational density of different tasks may not be equal. In order to simplify the calculation, the present invention uses the same computational density for the tasks of each TU.

[0129] In the TLOE framework, the present invention defines all available CPU resources as the set of each task j∈{1, 2, 3, ..., L, L+1, L+2}, where {1, 2, 3, ..., L} represents all available LEO satellite edge servers, L+1 represents the cloud server located on the ground, and L+2 represents the task directly computed on the local CPU. The binary computation offloading decision of the task of the lth TU calculated by the jth CPU is denoted as x ij .x i，L+1 1 means that the ith task is executed by the cloud server, x i，L+2 When it is 1, it indicates that the task is calculated locally, otherwise it is 0. ij When it is 1, the task is executed on the corresponding satellite edge server j.

[0130] The present invention assumes that each ground user can only access one LEO satellite for data transmission, and multiple ground users share the same spectrum resources, which means that there is mutual interference between ground users. Therefore, the uplink transmission rate from the i-th TU to the j-th satellite edge server can be calculated as:

[0131]

[0132] The channel power gain between the i-th TU and the j-th satellite edge server can be described as h ij The transmit power of the i-th TU can be expressed as σ 2 represents the noise power, B ij Represents the bandwidth allocated by the j-th LEO satellite to the i-th TU.

[0133] Local computing model: Due to the limited computing resources and power of mobile devices, the device allocates basic tasks to perform calculations locally as much as possible; the time spent on the local CPU processing of the task of the i-th TU is expressed as:

[0134]

[0135] in, is the CPU clock frequency of the i-th TU. The execution power of the i-th TU in local task calculation can be expressed as:

[0136]

[0137] k i represents the effective switch capacitance. V is a positive constant, generally set to 3. Therefore, the corresponding energy consumption of the i-th TU using the local calculation method is:

[0138]

[0139] 2) Satellite edge computing model: When using satellite edge computing methods for offloading calculations, the total processing delay should include three parts, namely, the transmission delay of the transmission task data Propagation delay at the speed of light Computational latency of tasks on satellite edge servers Although the satellite of the present invention is a low-orbit satellite, its distance from the ground is much greater than the distance between the TUs located on the ground. For simplicity, in this TLOE model, the present invention does not consider the inclination problem between the satellite and the ground TU. The present invention assumes that the actual distance between the TU and the satellite is consistent with the height of the LEO satellite above the ground, which is set to a constant H. The specific satellite-ground geometry model will be further studied in future work. If the i-th TU needs to offload the task to the satellite edge server j, the total processing time is:

[0140]

[0141] C is the speed of light, and is the CPU clock rate of the jth satellite edge server used to calculate the i-th TU task. The energy consumption of the low-orbit satellite offloading calculation process can be described as:

[0142]

[0143] 3) Cloud computing model: The central cloud server is located on the earth, and its computing power is much greater than that of the satellite edge server and the local CPU. Before the cloud server provides computing services to users, it needs to forward user tasks to the cloud server through satellite relay.

[0144] Similar to the delay model and energy consumption model of the LEO satellite edge server, the task calculation of the cloud server only has one more process of forwarding from the satellite to the cloud than the task calculation of the satellite edge server. The forwarding delay is expressed as It is expressed as the third term in (7). Similarly, in addition to ignoring the downlink propagation delay, the present invention also ignores the return process of the download task, so the delay model can be expressed as:

[0145]

[0146] The energy consumption model in cloud computing is:

[0147]

[0148] Where R EC is the uplink transmission rate from the edge server to the cloud server. C represents the CPU computing rate of the cloud server. The transmission power of the satellite edge server is P EC Therefore, for the task of the i-th TU, the total delay of task calculation can be expressed as:

[0149]

[0150] The total energy consumed by this task can be obtained from (10).

[0151]

[0152] 2.2 Problem Formulation

[0153] According to the above system model, the present invention can express the total cost of task processing as the weighted sum of energy consumption and delay when processing all tasks for the TLOE calculation system, which is expressed as the following cost function:

[0154]

[0155] The function The variables satisfy

[0156] Weight parameter satisfy The cost function needs to be adjusted according to the specific requirements of the TU. Specifically, for latency-sensitive tasks, If the user is more concerned with reducing the total energy consumption of the system, In extreme cases, when When , only the minimum total delay of TU is considered. When , the delay of TU is ignored and only the optimal energy consumption is considered.

[0157] Then the present invention can formulate an optimization problem and use the CNDO algorithm proposed by the present invention to solve it. The present invention has the following optimization problem:

[0158]

[0159]

[0160]

[0161]

[0162] x ij ∈{0,1} (12e)

[0163] Formulas (12b) and (12c) are bandwidth constraints, indicating that the allocated bandwidth must be non-negative. (12c) indicates that the total bandwidth allocated to all users must be less than the maximum bandwidth of the satellite edge server. For a TU task, only one CPU (including the local CPU of the TU) can be used, as shown in formula (12d). It should also be noted that the offload variable is a binary integer variable, as shown in formula (12e), and the definition is as described in Section 2. Formula (12) is a high-dimensional MIP problem, which is generally non-convex and belongs to an NP-hard problem, and is difficult to solve using traditional heuristic search algorithms. Therefore, the present invention will introduce a distributed deep learning algorithm to solve this problem in two stages.

[0164] 3. CNDO algorithm

[0165] In this section, the present invention will introduce a distributed algorithm CNDO for the cloud-edge-local three-layer architecture proposed above. The present invention uses multiple parallel DNN structures to speed up the calculation speed and convergence speed. The algorithm of the present invention divides the MIP problem (12) to solve it.

[0166] The main steps of CNDO architecture are as follows: Figure 4 As shown, in the first stage, S suboptimal decisions are obtained. In the second stage, these suboptimal decisions are used as known quantities to obtain the optimal decision by solving the bandwidth allocation problem in the original optimization problem, and the data pool data is updated at the same time. As the input of CNDO, it outputs the optimal offloading decision x after the network runs * The results of each operation are combined and stored in the data pool. When the data pool is full, the old data is continuously overwritten with updated data, and the DNN is trained to solve the above MIP problem (12).

[0167] 3.1 Generation of Unloading Decision Options

[0168] In the model of the present invention, the original unloading decision space is too large, that is, x * ∈{0, 1} N(L+2) In this method, the goal of the present invention is to obtain a suitable offloading strategy function π, which generates the current optimal offloading action of each DNN module in the entire distributed structure, and the current optimal offloading actions of all DNN modules constitute S suboptimal offloading actions of the entire network. As a reward to the network, the input task computation amount is the state. The calculation process is performed once in each time slot t, and there are: t∈{1, 2, ..., T}. The decision obtained by each DNN network is the optimal offloading action under this DNN. Using S DNN modules with the same structure, Although the S DNN modules have the same structure, the network parameters they update during training are Different. Its structure is as follows Figure 4 The first stage is shown.

[0169] 3.2 Resource allocation and updating data pool

[0170] The second phase of the algorithm includes resource allocation, optimal action generation, and data pool update. Resource allocation is to allocate available bandwidth to different users so that the weighted cost Minimum.

[0171] When these candidate binary offloading decisions are obtained through the calculation in the first stage, the present invention can consider that the offloading decision space has been expanded from 2 N(L+2) Reduced to S. At this stage, the present invention needs to select one action from these S actions that satisfies the minimum unloading cost As the final optimal action x *For each suboptimal offloading decision, it is considered as a known quantity in the original problem (12). Now we only need to solve the resource allocation variable The problem now is a convex optimization problem. The CVXPY tool can be used to calculate resource allocation The result is that the weighted cost It can be obtained by formula (12).

[0172] Next, the network input and the optimal offloading decision are concatenated as and stored in the data pool. The data pool is initially empty and has a size of μ. At the beginning, the network randomly generates parameters to complete the initialization. As the time slot t increases, the network will continue to fill When the data pool is full, the oldest data will be discarded and replaced by new data. All DNN blocks can randomly select several items in the data pool to train the network. Here, cross entropy is used as the loss function, and other loss functions can also be selected according to the specific model.

[0173] The following Algorithm 1 gives the CNDO algorithm process of the above-mentioned cloud-edge-local three-layer architecture.

[0174]

[0175]

[0176] 4. Simulation results

[0177] In this section, the present invention introduces the experimental parameters related to the simulation and uses simulation to evaluate the performance of the proposed CNDO algorithm. In all simulations, the present invention sets up a scenario including 6 remote TUs, a ground cloud server and a satellite with a MEC server, and temporarily does not consider the inter-satellite link. For the parallel DNN structure in the network, the present invention sets the epoch to 1000, and each DNN contains two hidden layers, using 100 and 80 neurons respectively. In the simulation, unless otherwise specified, the parameters of the proposed mode TLOE are summarized in Table 1. In order to take into account the optimization of both latency and energy consumption, the present invention weights and costs The weight parameter Designed to be 0.5.

[0178] The other parameters in the experiment are explained as follows: The total time frame T used in the present invention is 12000. The learning rate of the network is 0.01. The data pool size is set to 10 3 , the batch size is set to 128, and the training data accounts for 70% of the total data.

[0179] Since there are S parallel DNN structures in the algorithm of the present invention, the present invention explores the impact of the number of DNN structures on the normalization gain.

[0180] Table 1 Evaluation parameters

[0181]

[0182] Figure 5 The convergence performance of the six DNNs is shown. Figure 5 It can be seen that although the DNN structure is the same, the convergence curves of the six DNNs are slightly different. This is because the network parameters of the six DNNs are different. However, after 400 steps of training, the six curves tend to converge. This proves the effectiveness of the network and also shows that using a parallel DNN structure to generate different offloading strategies helps to diversify the offloading strategies.

[0183] To illustrate the performance of the proposed distributed algorithm, we compare the proposed method with the following baseline methods: i) local computing method: all local CPUs are used for the current task; ii) MEC method: all tasks are processed on the MEC server on the satellite side; iii) cloud computing method: all users use the cloud CPU for task computation.

[0184] The normalized weighted total cost comparison bar chart of the three offloading methods and the CNDO algorithm is shown in Figure 2. Figure 6 As shown. In the experiment of the present invention, 6 TUs and 6 DNNs are used for simulation. Here, the present invention normalizes the use of all methods according to the results of the local method for comparison, which is more intuitive and effective. Figure 6 It can be seen that if all vehicles use the local method, the normalized total cost is the highest, almost twice that of the cloud method and the CNDO algorithm. The cost of the MEC method is in the middle. The consumption of the CNDO method is slightly less than that of the cloud method. The performance of the CNDO method and the cloud method is similar. Because the number of users used in the present invention is not large, the cloud computing center can still bear the computational workload generated when all tasks are offloaded to the cloud CPU. In general, the model of the present invention has a lower total cost than other strategies. It further verifies the effectiveness of the CNDO method in the satellite-ground three-layer computing offloading architecture.

[0185] The present invention proposes a CNDO algorithm based on a three-layer cloud-edge-local satellite-assisted edge computing architecture to minimize the weighted cost of system energy consumption and latency. Using a distributed network composed of multiple parallel DNNs to make offloading decisions can effectively reduce the search space of offloading strategies and avoid the problem of dimensionality curse. The difficult MLP problem is reduced to a limited number of offloading decision variables, and the final calculation results are obtained in two stages. Simulation experiments show that the proposed strategy can produce better offloading decisions, verifying the good performance of the system.

[0186] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When the use is implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL) or wireless (e.g., infrared, wireless, microwave, etc.) mode) to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk Solid State Disk (SSD)), etc.

[0187] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for offloading computation of distributed deep learning in a satellite-ground collaborative network, characterized in that: The method for offloading calculation of distributed deep learning in the satellite-ground collaborative network comprises the following steps: Step 1: System model construction: Design of cloud-edge-local three-layer network TLOE for on-orbit edge mission computing. Tasks can be selectively executed directly locally or offloaded to the low-orbit satellite edge SatEC server or ground cloud computing center for execution. Step 2: Problem formulation: Model the offloading decision and resource allocation problem as a mixed integer programming (MIP) problem based on the task focus. Step 3: Calculate offloading: Construct a distributed offloading algorithm CNDO for the collaborative satellite-ground network to obtain the optimal offloading decision and resource allocation results; In step 1, the system model is constructed, including: The three-layer on-orbit edge computing network TLOE includes a central cloud server located on the ground, L LEO satellites and N ground user TUs, denoted as and Each LEO satellite is equipped with a MEC server to accept tasks from ground users and complete calculations. At the same time, the LEO satellite can also serve as a relay device for the cloud server to forward the task data of ground users. Each ground user has a task to be calculated in the current period. The model considers that when the current task of a user is calculated, the next task will be generated. All tasks are inseparable and can only be unloaded as a whole. The details of each task are grouped by variables {D i in , δ i } description, the two variables in the variable group represent the task input data size [in bits] and task processing density [in cycles / bit] of the i-th TU respectively; since the data volume of most task calculation results is small, the model does not consider the return process of the calculation results; therefore, the total number of CPU cycles required to complete the task calculation of the i-th TU is described as Z i =D i in δ i , where D i in and δ i The value of is obtained by analyzing the execution of each task; the computational density of different tasks may not be equal. To simplify the calculation, the same computational density can be used for the tasks of each TU; In the TLOE framework, all available CPU resources are defined as the set of each task j∈{1, 2, 3, ..., L, L+1, L+2}, where {1, 2, 3, ..., L} represents all available LEO satellite edge servers, L+1 represents the cloud server located on the ground, and L+2 represents the task directly computed on the local CPU; the binary computation offloading decision of the task of the i-th TU calculated by the j-th CPU is denoted as x ij ; If the ith user task is executed on the corresponding satellite edge server numbered j, then x ij is 1, otherwise 0; x i，L+1 1 means that the ith task is executed by the cloud server, x i，L+2 When it is 1, it means the task is calculated locally; Assume that each ground user can only access one LEO satellite for data transmission, multiple ground users share the same spectrum resources, and there is mutual interference between ground users; therefore, the uplink transmission rate from the i-th TU to the j-th satellite edge server is calculated as: Wherein, the channel power gain between the i-th TU and the j-th satellite edge server is described as h ij ; The transmit power of the i-th TU is expressed as σ 2 represents the noise power, B ij represents the bandwidth allocated to the i-th TU by the j-th LEO satellite; In step 2, the tasks focus on: According to the system model, for the cloud-edge-local three-layer architecture network, the execution of tasks should be carried out with emphasis for different task types or different application scenarios; among them, the emphasis includes the following two types: delay-sensitive and energy-sensitive; The delay-sensitive type focuses more on reducing the total delay of the system's computing offloading process; the energy-sensitive type focuses more on having lower total energy consumption during the computing offloading process; the task focus can be adjusted by setting weight parameters; In step 3, the construction of the collaborative satellite-ground network distributed offloading algorithm CNDO includes: In the CNDO architecture, S suboptimal decisions are obtained in the first stage. In the second stage, these suboptimal decisions are used as known quantities to obtain the optimal decision by solving the bandwidth allocation problem in the original optimization problem and completing the update of the data in the data pool; the CPU workload is As the input of CNDO, it outputs the optimal offloading decision x after the network runs * ; Merge the results of each operation and store them in a data pool. When the data pool is full, continuously overwrite the old data with updated data to train the DNN to solve the MIP problem; In step 3, the construction of the distributed offloading algorithm CNDO for the collaborative satellite-ground network also includes: (1) Generation of uninstall decision options In the model, the original unloading decision space is too large, that is, x * ∈{0, 1} N(L+2) ; Get a suitable offloading strategy function π, which generates the current optimal offloading action of each DNN module in the entire distributed structure. The current optimal offloading actions of all DNN modules constitute S suboptimal offloading actions of the entire network; subtract the total task processing cost from 1 As a reward to the network, the input task computation amount is the state; the calculation process is performed once in each time slot t, and there are: t∈{1, 2, ..., T}; the decision obtained by each DNN network is the optimal offloading action under this DNN; using S DNN modules with the same structure, S better offloading effects are obtained; although the S DNN modules have the same structure, the network parameters updated during the training process are different; (2) Resource allocation and data pool update The second stage of the algorithm includes resource allocation, optimal action generation and data pool update; resource allocation is to allocate available bandwidth to different users so that the weighted cost Minimum; When these candidate binary offloading decisions are obtained through the calculation in the first stage, it is considered that the offloading decision space has been expanded from 2 N(L +2) Reduce to S; you need to choose one of these S actions that satisfies the minimum unloading cost As the final optimal action x * Each suboptimal unloading decision is considered as a known quantity in the original problem; now only the resource allocation variables need to be solved The problem now is a convex optimization problem; the CVXPY tool is used to calculate resource allocation As a result, the weighted cost Obtained by the formula; The network input and optimal offloading decision will be concatenated as and stored in the data pool; the data pool is initially empty and has a size of μ; the network randomly generates parameters to complete initialization, and as the time slot t increases, the network will continue to fill When the data pool is full, the oldest data will be discarded and replaced by new data; all DNN blocks can randomly select several items in the data pool to train the network; cross entropy is used as the loss function here, and other loss functions can also be selected according to the specific model; The CNDO algorithm process of the cloud-edge-local three-layer architecture is as follows: 1) Input: CPU cycles of all TU tasks in each time frame 2) Output: The best offloading decision for each frame 3) Initialization: ① Initialize all DNN parameters with random parameters ② Clear the data pool of size μ; ③ For each time frame t, t=1, 2, ..., T, do the following operations ④ Input to SDNNs; ⑤ Generate candidate offloading decisions from the Sth DNN ⑥Solve the optimization problem and use Get resource allocation plan Compare the results and select the best one ⑦ If there is available space in the data pool, store the combined data to the data pool; ⑧Otherwise, overwrite the old data with the new data ⑨ Randomly extract S batches of data from the data pool to train DNN, and update network parameters during training 2. The method for offloading computation of distributed deep learning in a satellite-ground collaborative network as claimed in claim 1, characterized in that: In step 1, the system model construction further includes: (1) Local computing model: Due to the limited computing resources and power of mobile devices, the device allocates basic tasks to perform calculations locally as much as possible; the time spent on the local CPU processing of the task of the i-th TU is expressed as: in, is the local CPU clock frequency of the i-th TU; the execution power of the i-th TU in local task calculation is expressed as: Among them, k i represents the effective switch capacitance; V is a positive constant, set to 3; the corresponding energy consumption when the i-th TU adopts the local calculation method is: (2) Satellite edge computing model: When using satellite edge computing methods for offloading calculations, the total processing delay should include three parts: the transmission delay of the transmission task data. Propagation delay of light Computational latency of tasks on satellite edge servers Although the MEC server is set on a low-orbit satellite, the distance between it and the TU is much greater than the distance between the TUs on the ground. In this TLOE model, the inclination between the satellite and the ground TU is not considered. It is assumed that the actual distance between the TU and the satellite is consistent with the altitude of the LEO satellite over the ground, which is set as a constant H. If the i-th TU needs to offload the task to the satellite edge server j, the total processing time is: where C is the speed of light, and is the CPU clock rate of the jth satellite edge server used to calculate the i-th TU task; the energy consumption of the low-orbit satellite offloading calculation process is described as: (3) Cloud computing model: The central cloud server is located on the earth, and its computing power is much greater than that of the satellite edge server and the local CPU. Before the cloud server provides computing services to users, it needs to forward user tasks to the cloud server through satellite relay. Similar to the delay model and energy consumption model of the LEO satellite edge server, the task calculation process of the cloud server only has one more step of forwarding from the satellite to the cloud than the task calculation process of the satellite edge server; the forwarding delay is expressed as Similarly, in addition to ignoring the downlink propagation delay, the return process of the download task is also ignored, so the delay and energy consumption model is expressed as: The energy consumption of the cloud computing model is: Among them, R EC is the uplink transmission rate from edge server to cloud server, f C represents the CPU computing rate of the cloud server, and the transmission power of the satellite edge server is P EC ; For the task of the i-th TU, the total delay of task calculation is expressed as: The total energy consumed by the task is obtained from the following formula:

3. The method for offloading computation of distributed deep learning in a satellite-ground collaborative network as claimed in claim 1, characterized in that: In step 2, the problem is formulated, including: According to the system model, for the TLOE calculation system, the total cost of task processing is expressed as the weighted sum of energy consumption and delay when processing all tasks, expressed as the following cost function: Among them, the function The variables satisfy Weight parameter satisfy The cost function needs to be adjusted according to the specific requirements of TU; for delay-sensitive tasks, If users pay more attention to reducing the total energy consumption of the system, In extreme cases, when When , only the total delay of each TU is considered to be minimized; when When , the delay of TU is ignored and only the energy consumption is minimized; Formulate the optimization problem and use the proposed CNDO algorithm to solve it. There are the following optimization problems: x ij ∈{0，1}； Among them, the first two constraints are bandwidth limitations; Indicates that the allocated bandwidth must be a non-negative number; This means that the total bandwidth allocated to all users must be less than the maximum bandwidth of the satellite edge server; It means that for a task of a TU, only one CPU can be used, including the local CPU of the TU; the offload variable is a binary integer variable. The optimization problem is a MIP problem in a high-dimensional state, which is non-convex and NP-hard. It is difficult to solve it using a traditional heuristic search algorithm. Therefore, a distributed deep learning algorithm is introduced to solve this problem in two stages.

4. A distributed deep learning computing offloading system in a satellite-ground collaborative network implementing the distributed deep learning computing offloading method in a satellite-ground collaborative network as described in any one of claims 1 to 3, characterized in that: The computing offloading system for distributed deep learning in the satellite-ground collaborative network includes: System model building module for designing cloud-edge-local three-layer architecture network TLOE for on-orbit edge mission computing, which is executed locally, on SatEC servers or in ground cloud computing centers; Problem formulation module, which is used to model the offloading decision and resource allocation problem as a MIP problem; The computational offloading module is used to obtain the optimal offloading decision and resource allocation results by constructing a distributed offloading algorithm CNDO for the collaborative satellite-ground network.

5. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the computational offloading method for distributed deep learning in a satellite-ground collaborative network as claimed in claim 1.

6. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the computing offloading system for distributed deep learning in the satellite-ground collaborative network as described in claim 4.

Images