CN114153572A

CN114153572A - Calculation unloading method for distributed deep learning in satellite-ground cooperative network

Info

Publication number: CN114153572A
Application number: CN202111258473.8A
Authority: CN
Inventors: 陈晨; 李昊菲; 何辞; 张亚生; 刘雷
Original assignee: Xidian University; CETC 54 Research Institute
Current assignee: Xidian University; CETC 54 Research Institute
Priority date: 2021-10-27
Filing date: 2021-10-27
Publication date: 2022-03-08

Abstract

The invention belongs to the technical field of communication, and discloses a calculation unloading method for distributed deep learning in a satellite-ground cooperative network, which comprises the following steps: designing a cloud-edge-local three-layer architecture for an on-orbit edge computing scene, wherein tasks of a ground user can be executed in a local and low-orbit satellite edge server or a ground cloud computing center; modeling the unloading decision and resource allocation problem as a mixed integer programming problem; and constructing a distributed unloading algorithm aiming at the satellite-ground cooperative network to solve the mixed integer programming problem step by step and obtain an optimal unloading decision and resource allocation result. The problem addressed by the present invention is to minimize the latency energy consumption weighted total cost in the architecture subject to the constraints being met. In order to solve the problems, the invention provides a CNDO algorithm which obtains the optimal unloading decision in the TLOE scene in two steps, thereby avoiding the problem of dimension disaster. Experimental results prove that the scheme provided by the invention has good performance.

Description

Calculation unloading method for distributed deep learning in satellite-ground cooperative network

Technical Field

The invention belongs to the technical field of communication, and particularly relates to a calculation unloading method for distributed deep learning in a satellite-ground cooperative network.

Background

With the intensive research of the 6G technology, the traditional communication industry structure is undergoing a gradual and deep revolution. Due to the popularization of the 5G network and the Internet of things equipment, the demands of global users on emerging applications such as real-time transmission of high-definition videos, virtual reality, automatic driving and the like are greatly increased. Communication networks are gradually evolving towards the vision of user global coverage, fast response, anytime and anywhere. The flooding of large resource requests into the network presents new challenges to the limited distribution of computing resources in traditional communications.

Due to the regional limitations of complex terrains such as mountainous areas and deserts and remote areas such as rural areas, the traditional communication equipment is difficult to meet the QoS requirements of intelligent equipment in all areas. Under the influence of natural disasters, the ground communication network is easy to cause large-scale communication interruption due to damage of facilities. Meanwhile, the existing cloud computing center is far away from the mobile terminal, and the time delay requirement of emerging applications is difficult to meet.

The emergence of some emerging technologies gradually breaks the above dilemma. The appearance of a Mobile Edge Computing (MEC) network directly sinks Computing cache resources to the Edge of a user, and breaks the limit of the distance between a cloud server and the user. A large satellite constellation consisting of a plurality of Low Earth Orbit (LEO) satellites can meet the requirement of global area coverage, and has the characteristics of Low communication cost, small equipment size, small transmission delay and the like, and the propagation delay can reach millisecond level. The satellite-ground communication system derived from the low-earth orbit satellite integrates the advantages of ground communication facilities and a satellite network, can supplement and expand the ground network, and makes up for various limitations of the existing facilities.

Satellite communication is currently developing towards the direction of network broadband, coverage densification, application intelligence and world linkage. Similar to the computational offload optimization strategy in vehicle-mounted edge computing or drone assistance. For example, "y.wang, j.zhang, x.zhang, p.wang and l.liu," a Computing off-flow training in Satellite telecommunications Networks with Double Edge Computing, "2018 IEEE International Conference on Communication Systems (ICCS), pp.450-455,2018," and "j.zhang, x.zhang, p.wang, l.liu and y.wang," Double-Edge integrated Satellite telecommunications Networks, "in chip Communications, vol.17, vol.9, pp.128-146, sept.2020," provides a model of a Double-Edge star fusion network, and MEC-type servers are deployed on top of ground base stations and LEO satellites to ensure various requirements including Computing density and latency sensitivity. In "Y.Wang, J.Zhang, X.Zhang, P.Wang and L.Liu," A computer off-flow Stratagy in software project Networks with Double Edge Computing, "2018 IEEE International Conference on Communication Systems (ICCS), pp.450-455,2018," Hungarian algorithm is used to solve the problem of computational Offloading. "J.Zhang, X.Zhang, P.Wang, L.Liu and Y.Wang," Double-edge integrated satellite communication networks, "in China Communications, vol.17, No.9, pp.128-146, Sept.2020" mainly considers the caching and delivery problems of the satellite-to-ground network terminals, and adopts DQN to perform modeling, thereby reducing the average service delay of the system. The existing Satellite-ground network Computing unloading strategy based on the Game theory is provided in Y.Wang, J.Yang, X.Guo and Z.Qu, A Game-the electronic Approach to Computing off-filling in Satellite Edge Computing, in IEEE Access, vol.8, pp.12510-12520,2020. The above document considers a local-satellite two-layer model, and ignores the abundant computing resources of the ground cloud computing center. "Q.Tang, Z.Fei, B.Li and Z.Han," Computing off-filling in LEO Satellite Networks With Hybrid Cloud and Edge-filling, "in IEEE Internet of things Journal, vol.8, No.11, pp.9164-9176, Jul.2021" takes into account the problem of Satellite-to-ground coverage restrictions. It constructs a three-layer model that takes into account the satellite-aided case and provides a solution to minimize energy consumption using the ADMM algorithm.

On-track edge computing presents some difficulties while providing a convenient computing solution for end users. First, the limited channel capacity of the link will result in a non-negligible transmission delay between the satellite and the ground. Therefore, the problem of finding a reasonable bandwidth allocation and an optimal offloading decision is urgently needed to be solved. Second, the design offload decision and resource allocation problem is a Mixed-Integer Programming (MIP) problem. Conventional methods such as dynamic planning and branch-and-bound algorithms are not suitable for MEC networks with a large number of users. The traditional methods such as the game theory, the Hungarian algorithm and the like are adopted to obtain the optimal result, a large amount of iteration is needed, and the low time delay requirement of ground users such as vehicles and the like is difficult to meet. The potential and feasibility of deep learning in land edge computing provides a new direction for research.

With the rapid development of the Internet of Things (IoT) and its derived technologies such as the Internet of Things (Aerial Internet of Things, AIoT), the number of smart devices and their computing requirements have increased explosively. This makes it difficult for existing terrestrial base stations to guarantee the computational requirements and low latency requirements of the users. The edge computing architecture of an orbiting satellite provides a viable solution to the above challenges. For the existing on-orbit edge computing system, the following problems to be solved also exist: when the ground user equipment unloads the tasks, if the task size is not considered, the tasks are all unloaded to the edge equipment, which causes the congestion of the edge CPU request and the waste of local computing resources. Meanwhile, abundant computing resources of the ground cloud computing center are also ignored. And the task unloading of the ground user is a dynamic process, the traditional method mostly solves the problem of optimization at a certain moment at present, does not consider the research of the dynamic calculation unloading process in the continuous time of the ground user, and is difficult to meet the actual application.

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

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

(2) The designed unloading decision and resource allocation problem is an MIP problem, and the traditional methods such as dynamic planning and branch-and-bound algorithm are not suitable for MEC networks with a large number of users.

(3) The traditional methods such as the game theory, the Hungarian algorithm and the like are adopted to obtain the optimal result, a large amount of iteration is needed, and the low time delay requirement of ground users such as vehicles and the like is difficult to meet.

(4) With the explosive increase of the number of intelligent devices and the calculation requirements thereof, the existing ground base station is difficult to ensure the calculation requirements and low time delay requirements of users.

(5) When the ground user equipment unloads the tasks, if the task size is not considered, the tasks are all unloaded to the edge equipment, which causes the congestion of the edge CPU request and the waste of local computing resources. Meanwhile, abundant computing resources of the ground cloud computing center are also ignored.

(6) Task unloading of a ground user is a dynamic process, most of the conventional methods solve the problem of optimization at a certain moment, do not consider research on a dynamic calculation unloading process of the ground user in continuous time, and are difficult to meet the actual application.

The difficulty in solving the above problems and defects is: all the problems have certain processing difficulty. Aiming at the problem (1), low time delay is an important requirement of a future communication system, and solving the problem of overlarge time delay is always a difficult problem for researchers to design the communication system. For more and more emerging applications, it is a significant challenge and difficult point to comply with application innovation to propose new solutions. (5) And (6) the bottleneck of the conventional scheme at present, and the difficulty that people need to overcome in the satellite-ground processing system to ensure that the conventional ground computing resources are not wasted and the task unloading is compliant with the dynamic process in practical application is needed. In summary, the above problems all have certain processing difficulties for researchers in the field. The present invention is intended to give a corresponding solution to the above problems. The significance of solving the problems and the defects is as follows: the above problems and deficiencies are important breakthrough points for building satellite-assisted on-orbit edge computing networks and driving them into practical applications. Today, 5G networks provide a wide range of communication services to people, the communication demands of users and the types of applications are increasing. Conventional terrestrial networks have been inadequate to compensate for the supply and demand gaps created by the large application demands. The satellite-ground communication technology assisted by the low earth orbit satellite is classified as an important cut-in point and a main enabling technology which step into the 6G era by the industry due to the characteristics of wide ground coverage and the like. At present, the biggest difference between the 6G communication technology and the 5G technology is that the 6G draws a macro blueprint of global seamless coverage on the basis of 5G. This goal is difficult to meet with existing ground facilities only. Therefore, the non-ground equipment assists the successful deployment of the communication network and has great significance for the human society to step into the 6G era.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a distributed deep learning calculation unloading method in a satellite-ground cooperative network.

The invention is realized in such a way that a computation unloading method for distributed deep learning in a satellite-ground cooperative network comprises the following steps:

step one, system model construction: designing a cloud-Edge-local Three-layer architecture network (TLOE) construction system model for on-orbit Edge task computing, wherein tasks are selectively and directly executed locally or can be unloaded to an LEO Satellite Edge (SatEC) Server or a ground cloud computing center for execution;

step two, problem formulation: modeling an unloading decision and resource allocation problem into a mixed integer programming problem according to a task side focus;

step three, calculating and unloading: and constructing a Distributed offload algorithm (CNDO) aiming at the Collaborative satellite-ground Network to obtain an optimal offload decision and resource allocation result.

The scheme is formed by three steps, namely step one to step three, and is an important link for implementing the scheme. The overall implementation result of the three provides a new solution for the existing satellite-ground cooperation network, and provides a scheme combining deep learning for the development of the field of on-orbit edge calculation.

Further, in step one, the system model building includes:

the constructed three-layer on-orbit edge computing network TLOE comprises a ground central cloud server, L LEO satellites and N ground Users (TUs), and is recorded as

And

each LEO satellite is provided with an MEC server and used for receiving tasks of ground users and completing calculation, and meanwhile, the LEO satellite can also be used as relay equipment of a cloud server to complete forwarding of ground user task data; each land user has a task to be computed during the current time period. And when the current task calculation of a certain user is completed, generating the next task. All tasks are not divisible and can only be offloaded as a whole.

Variable set for details of each task { D_i ⁱⁿ，δ_iDescription, two variables in the variable group respectively represent task input of ith TU [ in bit unit }]And task processing density [ in cycles/bit units]. Since most tasks have small data quantity of calculation results, the model does not consider the return process of the calculation results. Thus, the total number of CPU cycles required to complete the task computation for the ith TU is depicted as Z_i＝D_i ⁱⁿδ_iWherein D is_i ⁱⁿAnd delta_iThe value of (c) is obtained by analyzing the execution of each task. The computational density of different tasks may not be equal, and the same computational density may be used for the tasks of each TU for simplicity of computation.

In the TLOE framework, all available CPU resources are defined as a set j e {1, 2, 3., L, L +1, L +2} of each task, wherein {1, 2, 3., L } represents all available LEO satellite edge servers, L +1 represents a cloud server located on the ground, and L +2 represents a direct computing task at the local CPU. Calculated by the jth CPUThe binary computation offload decision of the task of the ith TU is recorded as x_ij. If the ith user task is performed on the corresponding satellite edge server numbered j, x_ijIs 1, otherwise is 0. x is the number of_i，L+1To 1 indicates that the ith task is being performed by the cloud server, x_i，L+2A value of 1 indicates that the task is locally computed.

Assuming that each terrestrial user can only access one LEO satellite for data transmission, multiple terrestrial users share the same spectrum resource, which means that there is mutual interference between terrestrial users. Therefore, the uplink transmission rate from the ith TU to the jth satellite edge server is calculated as:

wherein the channel power gain between the ith TU and the jth satellite edge server is described as h_ij(ii) a The transmit power of the ith TU is denoted as

σ²Representing the noise power, B_ijIndicating the bandwidth allocated to the ith TU by the jth LEO satellite.

Further, in the step one, the system model building further includes:

(1) the local calculation model is as follows: due to the limited computational resources and power of mobile devices, devices allocate basic tasks to perform computations locally as much as possible. The time spent by the task of the ith TU in the local CPU is expressed as:

wherein the content of the first and second substances,

is the local CPU clock frequency of the ith TU; the execution power calculated by the ith TU in the local task is represented as:

wherein k is_iRepresents the effective switched capacitance; v is a normal number and is set to be 3; when the ith TU adopts the local calculation method, the corresponding energy consumption is as follows:

(2) satellite edge calculation model: when the satellite edge calculation method is adopted for carrying out unloading calculation, the total processing time delay comprises three parts, namely the transmission time delay of transmission task data

Propagation delay of light velocity propagation

Computing time delay for tasks on satellite edge servers

Although the satellite is a low orbit satellite, the distance between the satellite and the TU is far larger than that between TUs on the ground. In this TLOE model, the tilt problem between the satellite and the ground TU is not considered. The actual distance of TU from the satellite is assumed to coincide with the altitude to the earth of the LEO satellite, set to a constant H. If the ith TU needs to offload the task to the satellite edge server J, the total processing time is:

wherein C is the speed of light, and

is the CPU clock rate of the jth satellite edge server used to compute the ith TU task. The energy consumption of the low-earth-orbit satellite offloading calculation process is described as:

(3) cloud computing model: the central cloud server is located on earth and has much more computing power than the satellite edge servers and the local CPU. Before providing computing services for users, the cloud server needs to relay user tasks to the cloud server through a satellite.

Similar to the time delay model and the energy consumption model of the LEO satellite edge server, the task computation of the cloud server only has one more process of forwarding from the satellite to the cloud than the task computation of the satellite edge server. The forward delay is expressed as

Besides 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 as follows:

wherein R is_ECFor the upstream transmission rate, f, from the edge server to the cloud server^CThe calculation rate of a cloud server CPU is shown, and the transmitting power of a satellite edge server is P^EC. Therefore, for the task of the ith TU, the total delay calculated by the task is represented as:

the total energy consumed by the task is obtained from the following equation:

further, in step two, the problem is formulated, including;

according to the system model, aiming at the TLOE computing system, the total cost of task processing is expressed as the weighted sum of energy consumption and time delay when all tasks are processed, and is expressed as the following cost function:

wherein the function

Is satisfied with

Weight parameter

Satisfy the requirement of

The cost function needs to be adjusted according to the specific requirements of the TU. For delay-sensitive tasks, let

If the user pays more attention to reducing the total energy consumption of the system, the method satisfies

In extreme cases, when

Only the minimum total delay of the TU is considered. When in use

And in time, only the optimal energy consumption is considered by neglecting the time delay of the TU.

An optimization problem is formulated and solved by using the proposed CNDO algorithm, and the optimization problem comprises the following steps:

x_ij∈{0，1}

wherein the first two constraints are bandwidth limitations,

indicating that the allocated bandwidth must be non-negative.

Meaning that the total bandwidth allocated to all users must be less than the maximum bandwidth of the satellite edge server.

Indicating that only one CPU can be used for a task of a TU, including a local CPU of the TU; the unloading variable is a binary integer variable, the optimization problem is a MIP problem in a high-dimensional state, the MIP problem is non-convex, belongs to an NP-hard problem, and is difficult to solve by using a traditional heuristic search algorithm, so that a distributed deep learning algorithm is introduced, and the MIP problem is solved in two stages.

Further, in step two, the task emphasis points include:

according to the system model, aiming at the cloud-edge-local three-layer architecture network provided by the invention, the execution of tasks should be performed with a side emphasis for different task types or different application scenes; wherein, the emphasis points include the following two: delay sensitive and energy consumption sensitive;

the time delay sensitive type focuses more on reducing the total time delay of the system in the calculation unloading process; the energy consumption sensitive type focuses more on the lower total energy consumption in the calculation unloading process; the adjustment of the task emphasis can be completed by setting the weight parameters.

Further, in the third step, the construction of the distributed offload algorithm CNDO for the cooperative satellite-ground network includes;

in the CNDO system structure, S suboptimal decisions are obtained in the first stage, the suboptimal decisions are used as known quantities in the second stage, the optimal decisions are obtained by solving the problem of bandwidth allocation in the original optimization problem, and meanwhile, the data of a data pool is updated; loading CPU with work

As input of CNDO, and output of optimal uninstalling decision x after network operation^*(ii) a And merging and storing the results of each operation in a data pool, continuously covering old data with updated data after the data pool is full, and training DNN to solve the MIP problem.

Further, in step three, the constructing of the distributed offload algorithm CNDO for the cooperative satellite-ground network further includes:

(1) generation of offload decision options

In the model, the original offload decision space is too large, i.e., x^*∈{0，1}^N(L+2) _；The invention aims to obtain a proper unloading strategy function pi, the function generates the current optimal unloading action of each DNN module in the whole distributed structure, and the current optimal unloading actions of all the DNN modules form S suboptimal unloading actions of the whole network. With 1 minus the total cost of task processing

In return for the network, the calculated amount of the input task

Is a state. The calculation process is performed once per time slot t and has: t ∈ {1, 2,..., T }. Decisions obtained per DNN network

Is the optimal unloading action under this DNN. With S DNN modules of the same construction,

s good unloading effects are obtained. Although the S DNN modules have the same structure, the network parameters updated in the training process

Different.

(2) Resource allocation and update data pool

The second phase of the algorithm includes resource allocation, optimal action generation, and data pool update. The resource allocation is to allocate the available bandwidth to different users, so as to weight the cost

And minimum.

When these candidate binary offload decisions are derived by the first stage of computation, the offload decision space is considered to have been taken from 2^N(L+2)Reducing to S; one of the S actions needs to be selected to satisfy the minimum unloading cost

As the final optimal action x^*. Each sub-optimal offload decision is considered a known quantity in the original problem. Only the resource allocation result needs to be solved

The current problem isConvex optimization problem. CVXPY tool for computing resource allocation

As a result of (1), the weighted cost

Obtained by the formula.

The input and optimal offload decisions of the network will be spliced into

And stored in a data pool. The data pool is initially empty, with size μ. The network randomly generates parameters to complete initialization, and the network will continue to fill with the increase of the time slot t

When the data pool is full, the oldest data will be discarded and new data will be replaced. All DNN blocks can randomly choose several items in the data pool to train the network. The cross entropy is used as a loss function, and other loss functions can be selected according to a specific model;

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

1) inputting: CPU cycle of all TU tasks per time frame

2) And (3) outputting: optimal offload decision per frame

3) Initialization:

firstly, initializing all DNN parameters by using random parameters

Emptying a data pool with the size of mu;

③ for each time frame T, T ═ 1, 2

Fourthly, input

To SDNNs;

generating candidate offload decision from S-th DNN

Solving optimization problem and using

Obtaining a resource allocation plan

Comparing the results to select the optimum

Seventhly, if available space exists in the data pool, storing the combined data

To the data pool;

if not, overwriting the old data with the new data

Ninthly, randomly extracting S batch data from the data pool to train DNN, and updating network parameters during training

Another object of the present invention is to provide a computation offloading system for distributed deep learning in a satellite-ground cooperative network, which applies the computation offloading method for distributed deep learning in a satellite-ground cooperative network, where the computation offloading system for distributed deep learning in a satellite-ground cooperative network includes:

the system model building module is used for designing a cloud-edge-local three-layer architecture network TLOE for on-orbit edge task computing and executing the TLOE in a local, SatEC server or a ground cloud computing center;

the problem formulation module is used for modeling the unloading decision and the resource allocation problem as an MIP problem;

and the calculation unloading module is used for obtaining an optimal unloading decision and resource allocation result by constructing a distributed unloading algorithm CNDO (CNDO) aiming at the cooperative satellite-ground network.

It is a further object of the invention to provide a computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of:

(1) constructing a system model: designing a cloud-edge-local three-layer architecture network TLOE for on-orbit edge task computing, and executing the TLOE in a local, SatEC server or a ground cloud computing center;

(2) problem formulation: modeling the unloading decision and resource allocation problem as an MIP problem;

(3) calculating and unloading: and constructing a distributed unloading algorithm CNDO (CNDO) aiming at the cooperative satellite-ground network to obtain an optimal unloading decision and a resource allocation result.

The invention also aims to provide an information data processing terminal, which is used for realizing the computing unloading system for the distributed deep learning in the satellite-ground cooperative network.

By combining all the technical schemes, the invention has the advantages and positive effects that: the invention provides a satellite-ground three-layer hybrid architecture in compliance with the development vision of a 6G network on the basis of a ground communication network. The architecture breaks through the limitation that the abundant computing resources of a cloud computing center are ignored or a satellite is only used as an auxiliary relay device in the existing satellite-ground network, overcomes the application bottleneck of satellite communication in the industry, creatively provides a link that satellite edge computing is used as a three-layer satellite-ground network, and provides a new research idea for the deployment of a satellite communication system in the existing communication system.

The computing unloading method for the distributed deep learning in the satellite-ground cooperative network provided by the invention designs a computing unloading strategy for a user so as to minimize the weighted sum of time delay and energy consumption in a cloud-edge-local three-layer mixed AIoT framework. The tasks can be calculated at local and satellite ends and can also be forwarded to a cloud computing center; a satellite-ground network cooperative distributed offload algorithm (CNDO) based on a parallel neural network is proposed, and a plurality of optimal offload decisions in the AIoT are given. The experimental results show that this scheme has better performance than the other comparative schemes.

The invention models the unloading decision and resource allocation problem as MIP problem, and provides a distributed unloading algorithm (CNDO) aiming at the cooperative satellite-ground network to solve the problem and obtain the calculation unloading decision. The present invention simulates the proposed model and compares it to other baseline solutions to demonstrate that the proposed solution has good performance in terms of computational performance.

The invention provides a CNDO algorithm based on a three-layer cloud-edge-local satellite-assisted edge computing architecture, which aims to minimize the weighted cost of system energy consumption and time delay. The unloading decision is made by using a distributed network consisting of a plurality of parallel DNNs, so that the search space of the unloading strategy can be effectively reduced, and the problem of dimension disaster is avoided. Aiming at the difficult MLP problem, a limited number of unloading decision variables are found through a deep learning algorithm, and a final calculation result is obtained in two stages. Simulation experiments show that the proposed strategy can generate better unloading decisions, and the good performance of the system is verified.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the embodiments of the present invention will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.

Fig. 1 is a flowchart of a computation offloading method for distributed deep learning in a satellite-ground cooperative network according to an embodiment of the present invention.

FIG. 2 is a block diagram of a computing offloading system for distributed deep learning in a satellite-ground collaboration network according to an embodiment of the present invention;

in the figure: 1. a system model construction module; 2. a problem formulation module; 3. and a calculation unloading module.

Fig. 3 is a schematic diagram of a collaborative task offloading framework for computing offloading according to an embodiment of the present invention.

Fig. 4 is a schematic structural diagram of a satellite-ground cooperative distributed offload (CNDO) algorithm provided in an embodiment of the present invention.

FIG. 5 is a graph of convergence for a distributed offload algorithm using 6DNNs provided by an embodiment of the present invention.

Fig. 6 is a graph illustrating a comparison of normalized performance of various methods provided by embodiments of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

Aiming at the problems in the prior art, the invention provides a computation unloading method for distributed deep learning in a satellite-ground cooperative network, which is described in detail below with reference to the accompanying drawings.

As shown in fig. 1, the computation offloading method for distributed deep learning in a satellite-ground cooperative network provided by the embodiment of the present invention includes the following steps:

s101, system model construction: designing a cloud-edge-local three-layer architecture network TLOE for on-orbit edge task computing, and executing the TLOE in a local, SatEC server or a ground cloud computing center;

s102, problem formulation: modeling the unloading decision and resource allocation problem as an MIP problem;

s103, calculating and unloading: and constructing a distributed unloading algorithm CNDO (CNDO) aiming at the cooperative satellite-ground network to obtain an optimal unloading decision and a resource allocation result.

As shown in fig. 2, a computing offloading system for distributed deep learning in a satellite-ground collaboration network according to an embodiment of the present invention includes:

the system model building module 1 is used for designing a cloud-edge-local three-layer architecture network TLOE for on-orbit edge task computing and executing the TLOE in a local, SatEC server or a ground cloud computing center;

the problem formulation module 2 is used for modeling the unloading decision and resource allocation problem as an MIP problem;

and the calculation unloading module 3 is used for obtaining an optimal unloading decision and resource allocation result by constructing a distributed unloading algorithm CNDO (CNDO) for the cooperative satellite-ground network.

The technical solution of the present invention is further described below with reference to specific examples.

1. With the rapid development of the internet of things (IoT) and its derived technologies, such as the internet of things over the air (AIoT), the number of smart devices and their computing demands have increased explosively. This makes it difficult for existing terrestrial base stations to guarantee the computational requirements and low latency requirements of the users. The edge computing architecture of an orbiting satellite provides a viable solution to the above challenges. The invention designs a computation offloading strategy for a user to minimize the weighted sum of time delay and energy consumption in a cloud-edge-local three-layer hybrid AIoT architecture. The tasks can be calculated at a local side and a satellite side, and can also be forwarded to a cloud computing center. Then, a satellite-to-ground network coordinated distributed offload algorithm (CNDO) based on a parallel neural network is proposed, giving a plurality of optimal offload decisions in AIoT. The experimental results show that this scheme has better performance than the other comparative schemes.

The invention designs a three-layer architecture network for on-orbit edge task calculation. As shown in fig. 3, the user may choose to perform the task locally, or forward the task to a satellite edge server or a cloud server, so as to complete the task remotely. In order to obtain the optimal unloading decision and resource allocation result, the invention adopts a calculation unloading algorithm based on distributed deep learning. The main contributions of this study are summarized below:

the present invention contemplates a cooperative cloud-edge-local three-tier on-track edge computing architecture (TLOE). Wherein the tasks are indivisible and can be performed locally, by a SatEC server or by a ground cloud computing center. The aim of the invention is to minimize the weighted total cost of energy consumption and latency of the whole system while jointly optimizing offloading decisions and resource allocation.

The invention models the unloading decision and resource allocation problem as MIP problem, and provides a distributed unloading algorithm (CNDO) aiming at the cooperative satellite-ground network to solve the problem and obtain the calculation unloading decision.

The present invention simulates the proposed model and compares it to other baseline solutions to demonstrate the superior performance of the proposed solution in terms of computational performance.

The rest of the invention is arranged as follows. In a second subsection, the present invention introduces a system model and formulates optimization problems. In a third subsection, proposed algorithms are introduced that generate optimal binary offload decisions based on distributed deep learning. Simulation results are given in the fourth subsection.

2. System model and problem formulation

The invention provides a collaborative task offloading framework. The device can choose to perform task computation locally, and can also completely unload the task to a cloud server or a LEO satellite edge server for computation.

2.1 System model

A three-tier in-orbit edge computing network TLOE as described in this invention is shown in FIG. 3, and includes a ground central cloud server, L LEO satellites, and N ground users (TU), which may be written as

And

respectively. Each LEO satellite is equipped with an MEC server which can receive the tasks of the ground users and complete the calculation. Meanwhile, the LEO satellite can also be used as relay equipment of the cloud server to finish the forwarding of the task data of the ground user. Each ground user has a task to be calculated in the current time period, and the model considers that when the current task calculation of a certain user is completed,and then the next task is generated. All tasks are indivisible and the tasks can only be offloaded as a whole.

The specific details of each task may be in the variable set { D }_i ⁱⁿ，δ_iDescription, two variables in the variable group are respectively expressed as the task input data size of the ith TU [ in units of bits }]And task processing density [ in cycles/bit units]. Since most tasks have small data quantity of calculation results, the model does not consider the return process of the calculation results. Thus, the total number of CPU cycles required to complete the task computation for the ith TU can be described as z_i＝D_i ⁱⁿδ_i. These compounds D_i ⁱⁿAnd delta_iThe value of (c) can be obtained by analyzing the execution of each task. The computational density of different tasks may not be equal. To simplify the computation, the present invention uses the same computation density for the task of each TU.

In the TLOE framework, the invention defines all available CPU resources as a set j e {1, 2, 3., L, L +1, L +2} for each task, where {1, 2, 3., L } represents all available LEO satellite edge servers, L +1 represents a cloud server located on the ground, and L +2 represents a direct computation task at the local CPU. The binary computation unloading decision of the task of the ith TU computed by the jth CPU is recorded as x_ij。x_i，L+1To 1 indicates that the ith task is being performed by the cloud server, x_i，L+2A value of 1 indicates that the task is locally computed, otherwise the value is 0. In addition to the two cases described above, x_ijAt 1, the task is executed on the corresponding satellite edge server j.

The invention assumes that each ground user can only access one LEO satellite for data transmission, and a plurality of ground users share the same frequency spectrum resource, which means that mutual interference exists among the ground users. Therefore, the uplink transmission rate from the ith TU to the jth satellite edge server can be calculated as:

wherein the channel power gain between the ith TU and the jth satellite edge server can be described as h_ij. The transmission power of the ith TU can be expressed as

The local calculation model is as follows: due to the limited computational resources and power of the mobile device, the device allocates basic tasks as much as possible to perform the calculations locally; the time spent by the task of the ith TU in the local CPU is expressed as:

wherein the content of the first and second substances,

is the CPU clock frequency of the ith TU. The execution power of the ith TU in the local task calculation can be expressed as:

k_irepresenting the effective switched capacitance. V is a normal number and is generally 3. Therefore, when the ith TU adopts the local calculation method, the corresponding energy consumption is as follows:

2) satellite edge calculation model: when the satellite edge calculation method is adopted for carrying out unloading calculation, the total processing time delay comprises three parts, namely the transmission time delay of transmission task data

Propagation delay of light velocity propagation

Computing time delay for tasks on satellite edge servers

Although the satellite of the present invention is a low earth orbit satellite, it is far from the ground by the distance between the two TUs. For simplicity, in this TLOE model, the present invention does not consider the problem of tilt between the satellite and the ground TU. The invention assumes that the actual distance between TU and satellite is consistent with the ground height of LEO satellite, and sets it as constant H. The specific geometric model of the star-earth will be further studied in the future. If the ith TU needs to offload the task to the satellite edge server j, the total processing time is:

c is the speed of light, and

is the CPU clock rate of the jth satellite edge server used to compute the ith TU task. The energy consumption of the low-earth satellite offloading computation process can be described as:

3) cloud computing model: the central cloud server is located on earth and has much more computing power than the satellite edge servers and the local CPU. Before providing computing services for users, the cloud server needs to relay user tasks to the cloud server through a satellite.

Denoted as the third term in (7). Similarly, besides 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:

the energy consumption model in the cloud computing situation is as follows:

wherein R is_ECThe uplink transmission rate from the edge server to the cloud server. f. of^CRepresenting the cloud server CPU computation rate. The transmitting power of the satellite edge server is P^EC. Thus, for the ith TU task, the total latency calculated by the task can be expressed as:

the total energy consumed by the task can be obtained from (10).

2.2 problem formulation

According to the system model, the total cost of task processing can be expressed as a weighted sum of energy consumption and time delay when all tasks are processed by aiming at the TLOE computing system, and the total cost is expressed as the following cost function:

wherein the function

Is satisfied with

Weight parameter

Satisfy the requirement of

The cost function needs to be adjusted according to the specific requirements of the TU. In particular, for delay sensitive tasks, it should be assumed

If the user pays more attention to reducing the total energy consumption of the system, the condition should be satisfied

In extreme cases, when

Only the minimum total delay of the TU is considered. When in use

The invention can then formulate an optimization problem and solve it using the CNDO algorithm proposed by the invention. The invention has the following optimization problems:

x_ij∈{0，1} (12e)

equations (12b) and (12c) are bandwidth constraints, indicating that the allocated bandwidth must be non-negative. (12c) Meaning that the total bandwidth allocated to all users must be less than the maximum bandwidth of the satellite edge server. For a task of one TU, only one CPU (including a local CPU of the TU) can be used, as shown in equation (12 d). It should also be noted that the unload variable is a binary integer variable, as shown in equation (12e), and is defined as described in section 2. Equation (12) is a MIP problem with a high-dimensional state, is generally non-convex, belongs to the NP-hard problem, and is difficult to solve with the conventional heuristic search algorithm. Therefore, the invention introduces a distributed deep learning algorithm, which solves the problem in two stages.

3. CNDO algorithm

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 computation and convergence. The algorithm of the present invention partitions the MIP problem (12) to solve.

The main steps of the CNDO architecture are as shown in fig. 4, S suboptimal decisions are obtained in the first stage, and the suboptimal decisions are used as known quantities in the second stage, so that the optimal decisions are obtained by solving the bandwidth allocation problem in the original optimization problem, and the data in the data pool is updated. The invention loads CPU work

As input of CNDO, and output of optimal uninstalling decision x after network operation^*. The results of each operation are merged and stored in a data pool, and after the data pool is full, old data is continuously covered by updated data, and DNN is trained to solve the MIP problem (12).

3.1 Generation of offload decision options

In the model of the present invention, the original offload decision space is too large, i.e., x^*∈{0，1}^N(L+2). In the method, the objective of the invention is to obtain a suitable offload policy function pi, which generates the current optimal offload actions of each DNN module in the whole distributed structure, and the current optimal offload actions of all DNN modules constitute S suboptimal offload actions of the whole network. The invention reduces the total cost of task processing by 1

In return for the network, the calculated amount of the input task

Different. The structure is shown in the first stage of fig. 4.

3.2 resource allocation and update data pools

And minimum.

When these are obtained by the first stage of calculationIn the case of candidate binary offload decision, the present invention may consider the offload decision space to have gone from 2^N(L+2)Is reduced to S. At this stage, the present invention needs to select one of the S actions that satisfies the minimum unload cost

As the final optimal action x^*. For each sub-optimal offload decision, it is treated as a known quantity in the original problem (12). Only the resource allocation variable needs to be solved at this time

The problem is now a convex optimization problem. CVXPY tool useful for computing resource allocation

As a result of (1), thus weighting the costs

Can be obtained from formula (12).

Next, the input and optimal offload decisions for the network will be stitched into

And stored in a data pool. The data pool is initially empty, with size μ. Initially, 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 new data will be replaced. All DNN blocks can randomly choose several items in the data pool to train the network. The cross entropy is used as a loss function, and other loss functions can be selected according to a specific model.

The following algorithm 1 gives the CNDO algorithm procedure of the cloud-edge-local three-layer architecture proposed above.

4. Simulation result

In this section, the present invention introduces experimental parameters related to the simulation and uses the simulation to evaluate the performance of the proposed CNDO algorithm. In all simulations, the invention sets a scenario comprising 6 remote TUs, a ground cloud server and a satellite with MEC server, temporarily without considering inter-satellite links. For a parallel DNN architecture in a network, the present invention sets epoch to 1000, each DNN containing two hidden layers, using 100 and 80 neurons respectively. In the simulations, the parameters of the proposed mode TLOE are summarized in Table 1, unless otherwise stated. In order to give consideration to the optimization of time delay and energy consumption, the invention weights and costs

Weight parameter of

The design is 0.5.

Other parameters in the experiment are explained below: the total time frame T used in the present invention is 12000. The learning rate of the network is 0.01. Data pool size set to 10³The batch size is set to 128 and the training data accounts for 70% of the total data.

Because S parallel DNN structures exist in the algorithm, the influence of the number of the DNN structures on the normalization gain is discussed.

TABLE 1 evaluation parameters

Fig. 5 shows the convergence performance of 6 DNNs. As can be seen from fig. 5, although the DNN structures are the same, the convergence curves of the 6DNNs are slightly different. This is because the network parameters of the 6DNNs are different. However, after 400 training steps, all 6 curves tend to converge. This demonstrates the effectiveness of the network, while also demonstrating that the use of parallel DNN structures to generate different offload policies contributes to the diversity of offload policies.

To illustrate the performance of the proposed distributed algorithm, the present invention compares this method to the following baseline method: i) the method of local computation: all local CPUs using the current task; ii) MEC process: all task processing is on the aspect of the satellite at the MEC server; iii) method of cloud computing: all users use the cloud CPU to perform task computing.

The normalized weighted total cost comparison histogram of the three methods used for the comparison calculation offload with the CNDO algorithm is shown in fig. 6. In the experiments of the present invention, simulations were performed using 6 TUs and 6 DNNs. Here, the present invention normalizes the use of all methods for comparison based on the results of the local methods, which is more intuitive and efficient. As can be seen from fig. 6, if all vehicles adopt 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 process is in an intermediate position. The consumption of the CNDO method is slightly less than the cloud method. The CNDO method and the cloud method have similar performance, and because the number of users used by the method is not large, the cloud computing center can still bear the calculated amount generated when all tasks are unloaded to the cloud CPU. Overall, the overall cost of the model of the invention is lower than that of other strategies. The effectiveness of the CNDO method in a three-layer satellite-ground computing unloading architecture is further verified.

The invention provides a CNDO algorithm based on a three-layer cloud-edge-local satellite-assisted edge computing architecture, which aims to minimize the weighted cost of system energy consumption and time delay. The unloading decision is made by using a distributed network consisting of a plurality of parallel DNNs, so that the search space of the unloading strategy can be effectively reduced, and the problem of dimension disaster is avoided. The difficult-to-solve MLP problem is reduced to a limited number of unloading decision variables, and a final calculation result is obtained in two stages. Simulation experiments show that the proposed strategy can generate better unloading decisions, and the good performance of the system is verified.

In the above embodiments, the implementation may be wholly or partially realized by software, hardware, firmware, or any combination thereof. When used in whole or in part, can be implemented in a computer program product that includes one or more computer instructions. When loaded or executed on a computer, cause the flow or functions according to embodiments of the invention to occur, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a network of computers, 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, for example, the computer instructions may be transmitted from one website site, computer, server, or data center to another website site, computer, server, or data center via wire (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL), or wireless (e.g., infrared, wireless, microwave, etc.)). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that includes one or more of the available media. The usable medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., Solid State Disk (SSD)), among others.

The above description is only for the purpose of illustrating the present invention and the appended claims are not to be construed as limiting the scope of the invention, which is intended to cover all modifications, equivalents and improvements that are within the spirit and scope of the invention as defined by the appended claims.

Claims

1. A computation unloading method for distributed deep learning in a satellite-ground cooperative network is characterized by comprising the following steps:

step one, system model construction: the method designs a cloud-edge-local three-layer architecture network TLOE for on-orbit edge task computing, and tasks are selectively and directly executed locally or can be unloaded to a low-orbit satellite edge SatEC server or a ground cloud computing center for execution;

step two, problem formulation: modeling an unloading decision and resource allocation problem into a mixed integer programming MIP problem according to a task side focus;

step three, calculating and unloading: and constructing a distributed unloading algorithm CNDO (CNDO) aiming at the cooperative satellite-ground network to obtain an optimal unloading decision and a resource allocation result.

2. The computation offloading method for distributed deep learning in a satellite-ground cooperative network as claimed in claim 1, wherein in step one, the system model building comprises:

the three-layer on-orbit edge computing network TLOE provided by the invention comprises a central cloud server positioned on the ground, L LEO satellites and N ground users TU which are marked as

And

each LEO satellite is provided with an MEC server for receiving the tasks of the ground users and completing the calculation; meanwhile, the LEO satellite can also be used as relay equipment of the cloud server to finish the forwarding of the task data of the ground user; each ground user has a task to be calculated in the current time period; the model considers that when the current task of a certain user is calculated, the next task is generated; all tasks are inseparable, and the tasks can only be unloaded as a whole;

variable set for details of each task { D_i ⁱⁿ，δ_iDescription, two variables in the variable group respectively represent the task input data size of the ith TU [ in bits }]And task processing density [ in cycles/bit units](ii) a Because the data volume of the calculation results of most tasks is small, the return process of the calculation results is not considered by the model; thus, the total number of CPU cycles required to complete the task computation for the ith TU is depicted as Z_i＝D_i ⁱⁿδ_iWherein D is_i ⁱⁿAnd delta_iThe value of (a) is obtained by analyzing the execution of each task; the calculation densities of different tasks may not be equal, and the same calculation density can be used for the task of each TU for simplifying the calculation;

in a TLOE framework, defining all available CPU resources as a set j epsilon {1, 2, 3., L, L +1, L +2} of each task, wherein {1, 2, 3., L } represents all available LEO satellite edge servers, L +1 represents a cloud server located on the ground, and L +2 represents a local CPU direct computing task; the binary computation unloading decision of the task of the ith TU computed by the jth CPU is recorded as x_ij(ii) a If the ith user task is performed on the corresponding satellite edge server numbered j, x_ijIs 1, otherwise is 0; x is the number of_i，L+1To 1 indicates that the ith task is being performed by the cloud server, x_i，L+2A 1 indicates that the task is locally computed;

assuming that each ground user can only access one LEO satellite for data transmission, a plurality of ground users share the same frequency spectrum resource, and the ground users have mutual interference; therefore, the uplink transmission rate from the ith TU to the jth satellite edge server is calculated as:

3. The computation offloading method for distributed deep learning in a satellite-ground cooperative network as claimed in claim 1, wherein in step one, the system model building further comprises:

(1) the local calculation model is as follows: due to the limited computational resources and power of the mobile device, the device allocates basic tasks as much as possible to perform the calculations locally; the time spent by the task of the ith TU in the local CPU is expressed as:

wherein the content of the first and second substances,

Propagation delay of light propagation

Computing time delay for tasks on satellite edge servers

Although the MEC server is located on low orbit satellites,but the distance between the TU and the base is far larger than that between TUs on the ground; in the TLOE model, the inclination problem between the satellite and the ground TU is not considered; assuming that the actual distance between the TU and the satellite is consistent with the ground height of the LEO satellite, and setting the distance as a constant H; if the ith TU needs to offload the task to the satellite edge server j, the total processing time is:

wherein C is the speed of light, and

is used for calculating the jth satellite edge server CPU clock rate of the ith TU task; the energy consumption of the low-earth-orbit satellite offloading calculation process is described as:

(3) cloud computing model: the central cloud server is positioned on the earth, and the computing power of the central cloud server is much larger than that of the satellite edge server and the local CPU; before providing computing service for a user, the cloud server needs to relay the user task to the cloud server through a satellite;

similar to the time delay model and the energy consumption model of the LEO satellite edge server, the task computing process of the cloud server is only one more step of forwarding from the satellite to the cloud than the task computing process of the satellite edge server; the forward delay is expressed as

Similarly, besides 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 as follows:

wherein R is_ECFor the upstream transmission rate, f, from the edge server to the cloud server^CThe calculation rate of a cloud server CPU is shown, and the transmitting power of a satellite edge server is P^EC(ii) a For the task of the ith TU, the total time delay calculated by the task is represented as:

the total energy consumed by the task is obtained from the following equation:

4. the computing offloading method for distributed deep learning in a satellite-ground cooperative network as claimed in claim 1, wherein in step two, the task-oriented points comprise:

according to the system model, aiming at the cloud-edge-local three-layer architecture network provided by the invention, the execution of tasks is performed with emphasis in different task types or different application scenes; wherein, the emphasis points include the following two: delay sensitive and energy consumption sensitive;

5. The method for computation offloading of distributed deep learning in a satellite-ground cooperative network as recited in claim 1, wherein in step two, the problem formulation comprises:

according to the system model, for the TLOE computing system, the total cost of task processing is expressed as a weighted sum of energy consumption and time delay when all tasks are processed, and is expressed as a cost function as follows:

wherein the function

Is satisfied with

Weight parameter

Satisfy the requirement of

The cost function needs to be adjusted according to the specific requirements of the TU; for delay-sensitive tasks, let

In extreme cases, when

Only the total time delay of each TU is considered to be minimum; when in use

Time delay of TU is ignored, and only minimum energy consumption is considered;

x_ij∈{0，1}；

wherein, the first two constraint conditions are bandwidth limitation;

indicating that the allocated bandwidth must be non-negative;

indicating that the total bandwidth allocated to all users must be less than the maximum bandwidth of the satellite edge server;

indicating that only one CPU can be used for a task of a TU, including a local CPU of the TU; the unloading variable is a binary integer variable, the optimization problem is a MIP problem of a high-dimensional state, is non-convex, belongs to an NP-hard problem, and is difficult to solve by using a traditional heuristic search algorithm, so that a distributed depth study is introducedThe algorithm is used for solving the problem in two stages.

6. The computation offload method for distributed deep learning in a satellite-ground cooperative network according to claim 1, wherein in step three, the construction of the cooperative satellite-ground network distributed offload algorithm CNDO comprises:

in the CNDO system structure, S suboptimal decisions are obtained in the first stage, the suboptimal decisions are used as known quantities in the second stage, the optimal decisions are obtained by solving the problem of bandwidth allocation in the original optimization problem, and meanwhile, the data in a data pool is updated; loading CPU with work

7. The computation offload method for distributed deep learning in a satellite-ground cooperative network according to claim 1, wherein in step three, the building of the distributed offload algorithm CNDO for the satellite-ground cooperative network further comprises:

(1) generation of offload decision options

In the model, the original offload decision space is too large, i.e., x^*∈{0，1}^N(L+2)(ii) a The method aims to obtain a proper unloading strategy function pi, the function generates the current optimal unloading action of each DNN module in the whole distributed structure, and the current optimal unloading actions of all the DNN modules form S suboptimal unloading actions of the whole network; with 1 minus the total cost of task processing

In return for the network, the calculated amount of the input task

Is in a state; the calculation process is performed once per time slot t and has: t ∈ {112,..., T }; decisions obtained per DNN network

Is the optimal unloading action under the DNN; with S DNN modules of the same construction,

s good unloading effects are obtained; although the S DNN modules have the same structure, the network parameters updated in the training process

Different;

(2) resource allocation and update data pool

The second stage of the algorithm comprises resource allocation, optimal action generation and data pool updating; the resource allocation is to allocate the available bandwidth to different users, so as to weight the cost

Minimum;

when these candidate binary offload decisions are derived by the first stage of computation, the offload decision space is considered to have been taken from 2^N(L ⁺²⁾Reducing to S; one of the S actions needs to be selected to satisfy the minimum unloading cost

As the final optimal action x^*(ii) a Each sub-optimal offload decision, taken as a known quantity in the original problem; only the resource allocation variable needs to be solved at this time

The present problem is a convex optimization problem; CVXPY tool for computing resource allocation

As a result of (1), the weighted cost

Obtaining the result by a formula;

the input and optimal offload decisions of the network will be spliced into

And stored in a data pool; the data pool is initially empty, with size μ; the network randomly generates parameters to complete initialization, and the network will continue to fill with the increase of the time slot t

When the data pool is full, the oldest data is discarded, and new data is replaced; all DNN blocks can randomly select a plurality of items in the data pool to train the network; the cross entropy is used as a loss function, and other loss functions can be selected according to a specific model;

1) inputting: CPU cycle of all TU tasks per time frame

2) And (3) outputting: optimal offload decision per frame

3) Initialization:

firstly, initializing all DNN parameters by using random parameters

Emptying a data pool with the size of mu;

③ for each time frame T, T ═ 1, 2

Fourthly, the medicine is used for transfusionInto

To SDNNs;

generating candidate offload decision from S-th DNN

Solving optimization problem and using

Obtaining a resource allocation plan

Comparing the results to select the optimum

To the data pool;

if not, overwriting the old data with the new data

8. A computation uninstalling system for distributed deep learning in a satellite-ground cooperative network, which implements the computation uninstalling method for distributed deep learning in the satellite-ground cooperative network according to any one of claims 1 to 6, wherein the computation uninstalling system for distributed deep learning in the satellite-ground cooperative network comprises:

9. A computer device, characterized in that the computer device comprises a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to carry out the steps of:

(1) constructing a system model: designing a cloud-edge-local three-layer architecture network TLOE for on-orbit edge task computing, wherein tasks are selectively and directly executed locally or can be unloaded to a SatEC server or a ground cloud computing center for execution;

10. An information data processing terminal, characterized in that the information data processing terminal is used for implementing a computing off-load system of distributed deep learning in a satellite-ground cooperative network according to claim 7.