Low earth orbit satellite edge calculation unloading method based on attention mechanism

Abstract

The invention relates to the technical field of low-earth-orbit satellite communication, and provides a low-earth-orbit satellite edge calculation unloading method based on an attention mechanism, which comprises the following steps of: 1, generating a 500000-group sample set; 2, establishing a low earth orbit satellite moving edge calculation unloading model, and constructing a low earth orbit satellite calculation unloading objective function and constraint; 3, defining a Markov decision process tetrad in reinforcement learning; 4, constructing an unloading strategy solving model based on the attention network; and 5, setting a gradient descent method training parameter and a loss function, and optimizing a network parameter. According to the method, the edge calculation unloading strategy is formulated for the low-orbit satellite through the attention network, the problem that algorithm convergence needs to be achieved by greatly increasing the number of algorithm iterations during large-scale user unloading decision making is solved, and the method has the advantages of being high in algorithm robustness and task planning speed; the edge unloading decision making time of large-scale users is shortened, and the resource utilization rate of the system is optimized.

Description

Technical Field

[0001] The present invention relates to the field of low-orbit satellite communication technology, and specifically to a low-orbit satellite edge computing offloading method based on an attention mechanism. Background Art

[0002] Satellite communication systems have been widely used in many fields due to their wide coverage of communication services, high transmission reliability and the characteristics of being unaffected by the ground environment. Among them, low-orbit satellites rely on the advantages of flexible networking, low transmission path loss, and low transmission delay. They can cooperate with ground systems to achieve more continuous and wide-area communication services, which is one of the important research directions for realizing "space-ground integration" in the future. The optimization of edge computing offloading of low-orbit satellites can greatly improve the system business processing efficiency and reduce energy consumption. It is one of the hot issues in the current low-orbit satellite communication field.

[0003] There are two main methods for edge offloading strategy planning: one is a heuristic optimization algorithm based on iterative optimization, and the other is a fast optimization algorithm based on reinforcement learning. Among them, the heuristic task allocation method has a long operation time, and it is difficult to meet the real-time requirements when facing edge offloading planning problems with a large number of users. The machine learning-based task planning algorithm, although the overall operation time of the method is short, has poor model robustness when dealing with offloading problems with different numbers of users and satellites. Summary of the invention

[0004] In view of the shortcomings of the existing technology, the present invention considers the multi-user task offloading scenario under the coverage of multiple low-orbit satellites, with the goal of minimizing the system's time and energy consumption. At the same time, combined with the users' differentiated computing task requirements and the different resource limitations of multiple satellites, a low-orbit satellite edge computing offloading method based on the attention mechanism is proposed to formulate offloading strategies for multiple users.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a low-orbit satellite edge computing offloading method based on attention mechanism, comprising the following steps: Step 1: Generate a sample set: Generate a sample set of 500,000 groups; Step 2: Establish the low-orbit satellite mobile edge computing offloading model and construct the low-orbit satellite computing offloading objective function and constraints: (2a) Establishing a mobile edge computing offloading model Assume that there are M low-orbit satellites and N users in the model. Users and low-orbit satellites can communicate directly, and data can be shared between satellites. The decision variables in the model are , when the variable value is 1, it is defined as user n offloading the computing task to satellite m for computing, and when the variable value is 0, it is defined as user n performing local computing; (2b) Constructing the objective function of low-orbit satellite computing offloading Objective function: Computation task time consumption of user n, computation energy consumption. Task offloading transmission time consumption between user m and satellite n, data transmission time consumption, computation time consumption, data transmission energy consumption, computation energy consumption; (2c) Constructing low-orbit satellite computation offloading constraints The constraints for offloading computation of low-orbit satellites are: the computation time of the user's task is less than the upper limit of the user's task time, the user's task cannot be divided, and a task can only select one low-orbit satellite offloading task; Step 3, define the Markov decision process quadruple in reinforcement learning: (3a) Constructing the state space for reinforcement learning The state space mainly includes: user's mission information, low-orbit satellite resource information and current allocation information; (3b) Constructing the reinforcement learning action space The actions defined for reinforcement learning are offloading decisions made for each user, i.e., the set of satellites and the own platform that can be used for computing task offloading; (3c) Constructing the reward function After all users make the offloading decision, the decision establishment calculation begins. The calculation cost is the user's local calculation cost and the calculation cost of the user offloading the task to satellite m; Step 4: Build an attention network-based offloading strategy solution model: (4a) Extracting initial features of user data Two fully connected neural networks are used to extract the initial features of user data and low-orbit satellite data respectively.

[0006] in, and Data features after data dimension unification for user n and low-orbit satellite m respectively. , , and They are all feature vectors of fully connected networks; (4b) Extracting high-dimensional features of user and LEO satellite data Two fully connected neural networks are used to extract high-dimensional features of user and satellite data

[0007]

[0008] in, , and They are different parameter matrices of the fully connected neural network; (4c) Use the attention mechanism to obtain the user’s uninstall strategy

[0009] in, Define the weight between the high-dimensional features of user data and the high-dimensional features of each low-orbit satellite. The weight of the user's high-dimensional features and its own weight is compared. The weight between the user and all satellites is compared with its own weight, and the one with the largest weight is selected as the user's offloading strategy. That is, if the weight of the high-dimensional features of m satellites and the user's high-dimensional features is the largest, then the satellite is selected as the user's offloading strategy. If the user's own weight is the largest, then the user performs the computing task locally. (4d) Loop through step 4 (c) until all users have completed their uninstallation decision; Step 5: Set the gradient descent training parameters and loss function to optimize the network parameters.

[0010] The present invention provides a low-orbit satellite edge computing offloading method based on an attention mechanism. It has the following beneficial effects: The attention network of the present invention formulates an edge computing offloading strategy for low-orbit satellites, which overcomes the problem that when dealing with large-scale user offloading decision-making problems, the number of algorithm iterations needs to be greatly increased to achieve algorithm convergence as the number of users increases. The present invention has the characteristics of strong algorithm robustness and fast task planning speed, which helps to reduce the edge offloading decision-making time of large-scale users and optimize the resource utilization of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0012] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0013] Example: Please see attached Figure 1 An embodiment of the present invention provides a low-orbit satellite edge computing offloading method based on an attention mechanism, comprising the following steps.

[0014] Step 1: Generate a sample set: Generate a sample set of 500,000 groups; in the randomly generated sample set, the low-orbit satellite altitude is 750km and the number of users is , the user transmission power is , the number of CPUs required for the calculation task is , user computing power The computing resources of low-orbit satellites are The size of the input computing task is , the delay range that users can accept .

[0015] Step 2: Establish the low-orbit satellite mobile edge computing offloading model and construct the low-orbit satellite computing offloading objective function and constraints: (2a) Establish a mobile edge computing offloading model: Assume that there are M low-orbit satellites and N users in the model. Users and low-orbit satellites can communicate directly, and data can be shared between satellites. The task of user n is ,in, The number of CPU cycles required to complete the user task, is the data size of the user task, is the user's delay requirement. The decision variables in the model are When the variable value is 1, it is defined as user n offloading the computing task to satellite m for computing. When the variable value is 0, it is defined as user n performing local computing.

[0016] (2b) Constructing the objective function of low-orbit satellite computation offloading; The computing task time consumption of user n is: ,in is the CPU frequency of user n.

[0017] The computational energy consumption of user n is: ,in is the switched capacitor.

[0018] The task offloading transmission time consumption between user m and satellite n is: ,in is the straight-line distance between user m and satellite n. C is the speed of light.

[0019] The data transmission time between user m and satellite n is: ,in is the uplink transmission rate.

[0020] The computing time consumption of user m on satellite n is ,in The computing resources allocated by satellite m to user n.

[0021] The energy consumption of data transmission between user m and satellite n is: ,in, is the transmission power of user n.

[0022] The computational energy consumption between user m and satellite n is: (2c) Constructing the computational offloading constraints for low-orbit satellites; The computational offloading constraints for low-orbit satellites are: The user's task calculation time is less than the user's task time limit; User tasks cannot be divided; A mission can only select one low-orbit satellite to unload a mission.

[0023] Step 3, define the Markov decision process quadruple in reinforcement learning: (3a) Constructing the state space for reinforcement learning The state space mainly includes: the user's task information (the size of the computing task, the maximum delay that the user can accept, the transmission power, the number of CPUs required for computing), the low-orbit satellite resource information and the current allocation information. The state space is expressed as: ,in, is the computing resource of satellite m, and mean is the mean calculation. After the mean calculation, the dimension of the state space does not change with the change of the number of users and the number of satellites.

[0024] (3b) Constructing the reinforcement learning action space The actions defined for reinforcement learning are offloading decisions made for each user, i.e., the set of satellites and the own platform that can be used for computing task offloading.

[0025] (3c) Constructing the reward function After all users make uninstall decisions, the decision establishment calculation begins. User local calculation cost: The computational cost of the user offloading the task to satellite m: Reinforcement learning training is to maximize the decision reward, so the reward function is the negative value of the computational cost.

[0026] Step 4: Construct an attention network-based offloading strategy solution model; (4a) Extracting initial features of user data and LEO satellite data Two fully connected neural networks are used to extract the initial features of user data and low-orbit satellite data respectively. The input dimensions of the neural network are user data dimension and satellite data dimension respectively, and the output dimensions are both 128.

[0027]

[0028] in, and Data features after data dimension unification for user n and low-orbit satellite m respectively. , , and are all feature vectors of the fully connected network.

[0029] (4b) Extracting high-dimensional features of user data and low-orbit satellite data Two fully connected neural networks are used to extract high-dimensional features of user data. The input dimension of the neural network is 128 and the output dimension is 256.

[0030]

[0031]

[0032] in, , and They are different parameter matrices of the fully connected neural network. (4c) Use the attention mechanism to obtain the user’s uninstall strategy

[0033] in, Define the weight between the high-dimensional features of user data and the high-dimensional features of each low-orbit satellite. The weights of the user's high-dimensional features and its own weight are compared, and the weights between the user and all satellites are compared with its own weight. The largest weight is selected as the user's offloading strategy. That is, if the weights of the m satellite's high-dimensional features and the user's high-dimensional features are the largest, then the satellite is selected as the user's offloading strategy. If the user's own weight is the largest, then the user performs the computing task locally.

[0034] (4d) Loop through step 4 (c) until all users have completed their uninstallation decision; Step 5: Set the gradient descent training parameters and loss function to optimize the network parameters.

[0035] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A low-orbit satellite edge computing offloading method based on attention mechanism, characterized in that: The following steps are involved: Step 1, generate a sample set: Generate a sample set of 500,000 groups; Step 2: Establish the low-orbit satellite mobile edge computing offloading model and construct the low-orbit satellite computing offloading objective function and constraints: (2a) Establishing a mobile edge computing offloading model Assume that there are M LEO satellites and N users in the model. Users and LEO satellites can communicate directly, and data is shared between satellites. The decision variable in the model is x n,m ={0,1}, when the variable value is 1, it is defined as user n offloading the computing task to satellite m for computing, and when the variable value is 0, it is defined as user n performing local computing; (2b) Constructing the objective function of low-orbit satellite computing offloading Objective function: computing task time consumption of user n, computing energy consumption, task offloading transmission time consumption between user m and satellite n, data transmission time consumption, computing time consumption, data transmission energy consumption, computing energy consumption; (2c) Constructing low-orbit satellite computation offloading constraints The constraints for offloading computation of low-orbit satellites are: the computation time of the user's task is less than the upper limit of the user's task time, the user's task cannot be divided, and a task can only select one low-orbit satellite offloading task; Step 3, define the Markov decision process quadruple in reinforcement learning: (3a) Constructing the state space of reinforcement learning The state space mainly includes: user's mission information, low-orbit satellite resource information and current allocation information; (3b) Constructing the reinforcement learning action space The actions defined for reinforcement learning are offloading decisions made for each user, i.e., the set of satellites and the own platform that can be used for computing task offloading; (3c) Constructing the reward function After all users make the offloading decision, the decision establishment calculation begins, which is the user's local calculation cost and the calculation cost of the user offloading the task to satellite m; Step 4: Build an attention network-based offloading strategy solution model: (4a) Extracting initial features of user data Two fully connected neural networks are used to extract the initial features of user data and low-orbit satellite data respectively. in, and These are the data features after the data dimensions are unified for user n and low-orbit satellite m, respectively. W1, b1, W2 and b2 are all feature vectors of the fully connected network; (4b) Extracting high-dimensional features of user and LEO satellite data Two fully connected neural networks are used to extract high-dimensional features of user and satellite data Among them, W q , W k and W v They are different parameter matrices of the fully connected neural network; (4c) Use the attention mechanism to obtain the user's uninstall strategy in, Define the weight between the high-dimensional features of user data and the high-dimensional features of each low-orbit satellite. The weight of the user's high-dimensional features and its own weight is compared. The weight between the user and all satellites is compared with its own weight, and the one with the largest weight is selected as the user's offloading strategy. That is, if the weight of the high-dimensional features of m satellites and the user's high-dimensional features is the largest, then the satellite is selected as the user's offloading strategy. If the user's own weight is the largest, then the user performs the computing task locally. (4d) Loop through step 4(c) until all users have completed their uninstallation decision; Step 5: Set the gradient descent training parameters and loss function to optimize the network parameters.

Images