Satellite-ground fusion network scheduling method, device, medium and equipment

Abstract

The present invention discloses a satellite-ground fusion network scheduling method, device, medium and equipment, and relates to the field of satellite-ground fusion network technology. It is applied to a satellite-ground fusion network including multiple ground users, a central server and multiple satellites; the ground user uses the satellite as an unloading unit to unload the model parameters to the central server, and each ground user performs communication optimization locally, first determines the local calculation time representation, performs a quantization model analysis on the local model of each ground user, determines the number of quantization bits corresponding to the weight parameter of each local model to determine the local quantization model parameter, and then determines the transmission time representation of each ground user to determine the total delay of federated graph learning, and then constructs and solves the federated graph optimization model, so that each ground user and each satellite perform federated graph learning under the working parameters corresponding to each optimized variable obtained by the solution to perform network scheduling. The present invention reduces the transmission delay and can obtain better unloading decisions.

Description

Technical Field

[0001] The present invention relates to the field of satellite-ground fusion network technology, and in particular to a satellite-ground fusion network scheduling method, device, medium and equipment. Background Art

[0002] Traditional terrestrial cellular networks have low coverage in remote areas, limited computing capacity and low robustness level. Based on this, satellite-ground fusion networks have been proposed and applied. Satellite-ground fusion networks can provide ground users with ubiquitous connectivity, global coverage and flexible deployment capabilities, which will optimize network communication resources and realize communication task scheduling.

[0003] The satellite-ground fusion network architecture can provide global coverage, huge transmission bandwidth and powerful relay forwarding capabilities for a large number of ground users by integrating ground communications and satellite communications. However, the existing satellite-ground fusion network nodes are dynamic and complex, and the corresponding federated directed graph structure has significant characteristics. First, each graph node lacks a fixed order and reference coordinates; second, the entity data in the satellite-ground fusion network has multimodal and complex characteristics, making it difficult to directly apply graph neural network methods to optimize latency.

[0004] In the existing technology, some studies have been conducted to allocate resources through greedy quantization. For the greedy quantization strategy, each ground user evenly distributes the number of quantization bits according to the number of computer bytes. Other studies have been conducted to allocate resources through deep deterministic policy gradients. The deep deterministic policy gradients include two traditional Actor neural networks and Critic neural networks. This solution will optimize the CPU cycle frequency, transmission bandwidth, and the number of quantization bits through the Actor neural network. At the same time, the Critic neural network will help the Actor network to adjust the model parameters.

[0005] However, the greedy quantization strategy allocates the number of quantization bits to each user evenly, which makes it difficult to find the optimal number of quantization bits for each user; and the deep deterministic policy gradient scheme increases the neural network training time and causes huge transmission delays because it collects a large amount of global state information. Summary of the invention

[0006] Based on this, it is necessary to provide a satellite-ground fusion network scheduling method, device, medium and equipment to address the above technical problems.

[0007] The present invention adopts the following technical solutions:

[0008] The present invention provides a satellite-ground fusion network scheduling method, which is applied to a satellite-ground fusion network including multiple ground users, a central server and multiple satellites; the ground user uses the satellite as an unloading unit, and unloads local model parameters to the central server through the satellite to perform federated graph learning of the satellite-ground fusion network scheduling; the method comprises:

[0009] Each ground user builds a local model of task processing and executes it locally based on the communication task and takes the CPU cycle frequency of each ground user as the optimization target; determines the local computing time representation of each ground user for executing the communication task, performs quantitative analysis on the weight parameters of the local model of each ground user, determines the number of quantization bits corresponding to the weight parameters of each local model to determine the quantized local quantization model parameters, and then determines the transmission time representation of each ground user, so as to determine the total delay of federated graph learning based on the local computing time representation and transmission time representation of each ground user; then, takes the CPU cycle frequency of each ground user, the transmission bandwidth between each ground user and the satellite, the transmission data volume of each ground user, and the transmission link graph between the ground user and the satellite as optimization variables, takes the total delay of federated graph learning as the optimization target, builds and solves the federated graph optimization model, and enables each ground user and each satellite to perform federated graph learning under the working parameters corresponding to the solved optimization variables for network scheduling.

[0010] The present invention provides a satellite-ground fusion network scheduling device, comprising:

[0011] A satellite-ground fusion network of multiple ground users, a central server and multiple satellites; the ground users use satellites as unloading units to unload local model parameters to the central server through the satellites to perform federated graph learning for satellite-ground fusion network scheduling;

[0012] Each ground user builds a local model for task processing and executes it locally based on the communication task and the CPU cycle frequency of each ground user as the optimization target;

[0013] The scheduling module is used to determine the local computing time representation of each ground user performing a communication task; perform quantitative analysis on the weight parameters of the local model of each ground user, determine the quantization bit number representation corresponding to the weight parameters of each local model to determine the quantized local quantization model parameters; determine the information transmission rate representation of the ground user unloading the local quantization model parameters according to the transmission link graph representation between the ground user and the satellite; determine the transmission time representation of each ground user according to the information transmission rate representation of each ground user; determine the total delay of the federated graph learning of the satellite-ground fusion network scheduling according to the local computing time representation and transmission time representation of each ground user performing the communication task; construct a federated graph optimization model with the CPU cycle frequency of each ground user, the transmission bandwidth between each ground user and the satellite, the quantization bit number of each ground user, and the transmission link graph between the ground user and the satellite as optimization variables, and minimize the total delay of the federated graph learning as the optimization goal; transmit and aggregate node information in the dynamic satellite-ground fusion network through spatial graph learning to establish a transmission link graph, solve the federated graph optimization model according to the established transmission link graph through deep federated reinforcement learning, and make each ground user and each satellite perform federated graph learning under the working parameters corresponding to each optimized variable obtained by solving to perform network scheduling.

[0014] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned satellite-ground fusion network scheduling method is implemented.

[0015] The present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned satellite-ground fusion network scheduling method when executing the program.

[0016] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0017] The present invention first determines the local computing time representation of the ground user performing the communication task, and then quantitatively analyzes the weight parameters of the local model of the ground user, thereby compressing the local model, reducing the amount of data unloaded by the ground user to the satellite in the graph optimization federated learning of the satellite-ground fusion network scheduling, and reducing the transmission delay; after quantitatively analyzing the local model, the transmission time representation of each ground user when performing the unloading is further determined, thereby determining the total delay of the graph optimization federated learning of the satellite-ground fusion network scheduling, and based on this, a federal graph optimization model of local computing, communication and quantization of the ground user is established, and solved through federal graph learning; wherein, the federal graph learning first processes the node information in the dynamic satellite-ground fusion network through spatial graph learning, and then further solves according to the result of the federal graph learning through deep federal reinforcement learning, and realizes efficient graph optimization federated learning of the dynamic satellite-ground fusion network by decomposing the two-step process, and can obtain better unloading decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0019] Figure 1 A schematic diagram of a satellite-ground fusion network scheduling method provided by the present invention;

[0020] Figure 2 A schematic diagram of a directed federated graph scenario of a satellite-ground fusion network provided by the present invention;

[0021] Figure 3 A schematic diagram of a federated graph learning and deep federated reinforcement learning framework provided by the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. 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 making creative work are within the scope of protection of the present invention.

[0023] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0024] Figure 1The present invention is a flow chart of a satellite-ground fusion network scheduling method, which is applied to a satellite-ground fusion network including multiple ground users, a central server and multiple satellites; the ground user uses the satellite as an unloading unit to unload the local model parameters to the central server through the satellite to perform federated graph learning for satellite-ground fusion network scheduling, which specifically includes the following steps:

[0025] S101: Each ground user builds a local model of task processing and executes it locally according to the communication task and takes the CPU cycle frequency of each ground user as the optimization target; and determines the local computing time representation of each ground user executing the communication task.

[0026] S102: Quantitatively analyze the weight parameters of the local model of each ground user, determine the number of quantization bits corresponding to the weight parameters of each local model to determine the quantized local quantization model parameters; determine the information transmission rate representation of the ground user unloading the local quantization model parameters based on the transmission link diagram between the ground user and the satellite; determine the transmission time representation of each ground user based on the information transmission rate representation of each ground user.

[0027] S103: Determine the total delay of federated graph learning for satellite-ground fusion network scheduling according to the local computing time representation and transmission time representation of each ground user performing the communication task.

[0028] S104: Construct a federated graph optimization model by taking the CPU cycle frequency of each ground user, the transmission bandwidth between each ground user and the satellite, the number of quantized bits of each ground user, and the transmission link graph between the ground user and the satellite as optimization variables, and taking the minimization of the total delay of federated graph learning as the optimization goal.

[0029] S105: The node information in the dynamic satellite-ground fusion network is transmitted and aggregated through spatial graph learning to establish a transmission link graph. The federated graph optimization model is solved according to the established transmission link graph through deep federated reinforcement learning, and each ground user and each satellite performs federated graph learning under the corresponding working parameters of each optimized variable obtained to perform network scheduling.

[0030] Figure 2 FIG. 1 is a schematic diagram of a directed federated graph scenario of a satellite-ground fusion network in the present invention. Figure 2 As shown, the present invention introduces a directed federated graph model of the satellite-ground fusion network, which can be represented as a node-edge pair ,in, is the set of nodes in the node-edge pair (including nodes corresponding to ground users and satellites), is a set of edges.

[0031] The set of ground users and satellites can be defined as and For the proposed directed federated graph model, each ground user Select the best satellite unit to offload local model parameters instead of associated private data, where each node Includes relevant local information such as transmit power or noise.

[0032] Furthermore, if the ground user Select Satellite As an unloading unit, an edge collection will be established Based on the above node and edge features, the node matrix is ​​defined as ,in is the total number of columns. Therefore, if , ;otherwise, ,in is the adjacency feature matrix The number of columns, Each ground user choose As an unloading unit, Is a user and satellite The channel gain between .

[0033] For the satellite-ground fusion network scenario, the present invention assumes that each ground user Generate communication tasks ,in It is task vectors, where Is a user Therefore, the total task size of all ground users is expressed as , each user The loss function of federated learning is expressed as , are the weight parameters of the neural network and is the model error function. Therefore, the global loss function of federated learning is expressed as: , .

[0034] Communication tasks refer to specific tasks such as compute-intensive tasks and delay-sensitive tasks. The local model of ground users can build a local model for task processing. The input of the local model includes the number of CPU cycles and task volume required to process one bit of task. The output includes local computing delay and energy consumption. The optimization target includes CPU cycle frequency.

[0035] In order to protect data privacy, the present invention considers each task to be executed by the local user. Because the satellite's transmission power is higher than that of the ground user and the task return amount is lower, the present invention ignores the feedback delay. Therefore, the local computing time of the task is: .

[0036] in, The number of CPU cycles required to process a 1-bit task for the nth ground user, is the CPU cycle frequency of the nth ground user. Meanwhile, the energy consumption generated by processing tasks is defined as: .in, Hardware parameters for executing the mission for the nth ground user.

[0037] Based on the above-mentioned federated learning model and local computing model, in order to reduce power consumption and improve computing speed, the present invention introduces a quantization model. First, the local model parameter update process is expressed as: , is the loss function after gradient calculation.

[0038] in, is the learning rate of parameter update. Therefore, the global weight parameter of federated graph learning is expressed as: , Learning global weight parameters for federated graphs.

[0039] Next, the present invention defines the weight parameter ,in, is the number of dimensions of the neural network. Pick the minimum value and maximum value , according to the sequence size , the number of quantization bits required is And the quantization interval is expressed as .

[0040] Furthermore, the quantized sequence can be expressed as , , ,in, .

[0041] in, When the weight parameter After falling into the quantization sequence, the relevant quantization function is represented as: .

[0042] in, , , .

[0043] After obtaining the quantization parameter, the total number of quantization bits is: .

[0044] Based on the above task scheduling model, the present invention introduces a federated graph model in the satellite-ground fusion network. Transmitting local quantization model parameters to a remote server, the present invention uses orthogonal frequency division multiple access to transmit tasks, therefore, The information transmission rate of the local quantization model parameters unloaded by the ground user is: .

[0045] in, is the transmit power, is the optimized transmission bandwidth, is the noise power. At the same time, On the ground user and remote satellite The channel gain between is the channel interference from other regions. Therefore, the total delay caused by local execution and transmission is: .

[0046] Next, when transmitting the model parameters to the remote satellite, the transmission energy consumption is defined as: .

[0047] Furthermore, the present invention defines the optimization variable Based on the optimization variables , the node matrix and the adjacency feature matrix , the federated graph optimization problem is defined as:

[0048] .

[0049] Among them, C1 and C2 indicate that the CPU cycle frequency and transmission power cannot exceed the maximum value and C3 represents each user The total energy consumption cannot exceed the maximum C4 means that the total number of quantization bits should be lower than the total number of transmission bits.

[0050] Figure 3 Schematic diagram of a federated graph learning and deep federated reinforcement learning framework in the present invention. Figure 3As shown, the present invention proposes a federated graph learning algorithm to optimize communication, computing and quantization resources, and decomposes the federated graph learning algorithm into spatial graph learning and deep federated reinforcement learning, which can obtain the optimal offloading decision. First, the spatial graph learning algorithm is explored to adapt to the dynamic graph structure and dynamic node arrangement. After outputting the graph type, the present invention proposes a new deep federated reinforcement learning algorithm to optimize the variables.

[0051] (1) Spatial graph learning:

[0052] Since multilayer perceptrons are highly sensitive to graph size and node arrangement, it is difficult to directly apply traditional methods to handle a variety of graph structures. Therefore, the present invention designs a new spatial graph learning method to transfer and aggregate node information, which can overcome dynamic node arrangement. Therefore, the convolution operation is defined as: .

[0053] Based on this prior knowledge, the spatial graph calculation can be expressed as: .

[0054] in, And D is the degree matrix of each node. The graph type between ground users and satellites in the satellite-ground fusion network. learning neural network activation functions for spatial graphs, is the node matrix, Learn the weights of the kth layer of a neural network for a spatial graph, Bias for the kth layer of the spatial graph learning neural network, is the adjacency feature matrix, is the convolution operation, is a graph node, is the convolution kernel, is the image pixel matrix, is the convolution kernel matrix.

[0055] Based on this, according to the graph type between ground users and satellites, deep federated reinforcement learning can be used, with the weight parameters of the local models of each ground user and the global weight parameters of the federated graph learning as the state, the optimized variables as the action, and the reward determined according to the total delay of the federated graph learning, so as to solve the federated graph optimization model.

[0056] (2) Deep Federated Reinforcement Learning:

[0057] Further, based on the above-mentioned federated graph learning, the present invention can obtain graph types to perform optimization variables. Specifically, when each ground user n offloads local model parameters to the remote server m, a transmission link graph is established, which is regarded as a satellite-user node pair. At the same time, when the distance is less than the threshold, each node pair will suffer from the neighboring channel interference, which is regarded as a neighboring graph link. In the model of the present invention, it is assumed that there are a total of satellite-user pairs.

[0058] Therefore, since the above model includes multiple integer variables and continuous variables, it is difficult to solve by traditional methods. Therefore, the present invention uses the Actor-Critic network structure to process the optimization variables, where the state, action and reward functions are as follows: State: .

[0059] in, is the global model weight parameter and is the weight parameter for each ground user.

[0060] Action: For each iteration round , each agent needs to optimize the CPU cycle frequency , transmission bandwidth and the number of quantization bits ,in .

[0061] Reward: The reward function is defined as .

[0062] based on Figure 1 In the satellite-ground fusion network scheduling method shown, each ground user builds a local model of task processing and executes it locally based on the communication task and takes the CPU cycle frequency of each ground user as the optimization target; determines the local computing time representation of each ground user performing the communication task, performs quantization model analysis on the local model of each ground user, determines the quantization bit number representation corresponding to the weight parameter of each local model, and then determines the transmission time representation of each ground user, so as to determine the total delay of federated graph learning based on the local computing time representation and transmission time representation of each ground user, and then takes the CPU cycle frequency of each ground user, the transmission bandwidth between each ground user and the satellite, the transmission data volume of each ground user, and the transmission link graph between the ground user and the satellite as optimization variables, takes the total delay of federated graph learning as the optimization target, builds a federated graph optimization model and solves it, so that each ground user and each satellite can perform federated graph learning under the working parameters corresponding to the solved optimization variables to perform network scheduling.

[0063] The present invention establishes a directed federated graph model for satellite-ground fusion network. Based on node characteristics and link characteristics, the present invention designs a graph convolution model to process node dynamics and multimodal characteristics. A federated graph learning algorithm is proposed to optimize communication, computing and quantization resources, and the federated graph learning algorithm is decomposed into spatial graph learning and deep federated reinforcement learning, which can obtain the optimal offloading decision.

[0064] When applying the satellite-ground fusion network scheduling method provided, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0065] The above is a satellite-ground fusion network scheduling method provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding satellite-ground fusion network scheduling device. It includes:

[0066] A satellite-ground fusion network of multiple ground users, a central server and multiple satellites; the ground users use satellites as unloading units to unload local model parameters to the central server through the satellites to perform federated graph learning for satellite-ground fusion network scheduling;

[0067] Each ground user builds a local model for task processing and executes it locally based on the communication task and the CPU cycle frequency of each ground user as the optimization target;

[0068] The scheduling module is used to determine the local computing time representation of each ground user performing a communication task; perform quantitative analysis on the weight parameters of the local model of each ground user, determine the quantization bit number representation corresponding to the weight parameters of each local model to determine the quantized local quantization model parameters; determine the information transmission rate representation of the ground user unloading the local quantization model parameters according to the transmission link graph representation between the ground user and the satellite; determine the transmission time representation of each ground user according to the information transmission rate representation of each ground user; determine the total delay of the federated graph learning of the satellite-ground fusion network scheduling according to the local computing time representation and transmission time representation of each ground user performing the communication task; construct a federated graph optimization model with the CPU cycle frequency of each ground user, the transmission bandwidth between each ground user and the satellite, the quantization bit number of each ground user, and the transmission link graph between the ground user and the satellite as optimization variables, and minimize the total delay of the federated graph learning as the optimization goal; transmit and aggregate node information in the dynamic satellite-ground fusion network through spatial graph learning to establish a transmission link graph, solve the federated graph optimization model according to the established transmission link graph through deep federated reinforcement learning, and make each ground user and each satellite perform federated graph learning under the working parameters corresponding to each optimized variable obtained by solving to perform network scheduling.

[0069] For the specific definition of the satellite-ground fusion network scheduling device, please refer to the definition of the satellite-ground fusion network scheduling method above, which will not be repeated here. Each module in the above-mentioned satellite-ground fusion network scheduling device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0070] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A satellite-ground fusion network scheduling method is provided.

[0071] The present invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 A satellite-ground fusion network scheduling method is provided.

[0072] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0073] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A satellite-ground fusion network scheduling method, characterized in that: The method is applied to a satellite-ground fusion network including multiple ground users, a central server and multiple satellites; the ground users use satellites as unloading units and unload local model parameters to the central server through the satellites to perform federated graph learning for satellite-ground fusion network scheduling; the method includes: Each ground user builds a local model of task processing and executes it locally based on the communication task and takes the CPU cycle frequency of each ground user as the optimization target; determines the local computing time representation of each ground user performing the communication task; The weight parameters of the local model of each ground user are quantitatively analyzed, and the quantization bit number expression corresponding to the weight parameter of each local model is determined to determine the quantized local quantization model parameter; according to the transmission link diagram between the ground user and the satellite, the information transmission rate expression of the ground user to unload the local quantization model parameter is determined; according to the information transmission rate expression of each ground user, the transmission time expression of each ground user is determined; According to the local computation time representation and transmission time representation of each ground user performing the communication task, the total delay of the federated graph learning for satellite-ground fusion network scheduling is determined; The CPU cycle frequency of each ground user, the transmission bandwidth between each ground user and the satellite, the number of quantized bits of each ground user, and the transmission link graph between the ground user and the satellite are used as optimization variables, and the minimum total delay of federated graph learning is used as the optimization goal to build a federated graph optimization model. Through spatial graph learning, the node information in the dynamic satellite-ground fusion network is transmitted and aggregated to establish a transmission link graph. Through deep federated reinforcement learning, the federated graph optimization model is solved according to the established transmission link graph, and the ground users and satellites perform federated graph learning under the corresponding working parameters of the solved optimization variables to perform network scheduling.

2. The satellite-ground fusion network scheduling method according to claim 1, characterized in that: The determining of the local computing time representation of each ground user performing the communication task specifically includes: The local computing time for each ground user to perform a task is determined by the following formula based on the CPU cycle frequency of each ground user: ; in, The local computation time to perform the communication task for the nth ground user, The number of CPU cycles required to process a 1-bit task for the nth ground user, For the communication task of the nth ground user, is the CPU cycle frequency of the nth ground user.

3. The satellite-ground fusion network scheduling method according to claim 1, characterized in that: The step of performing quantization analysis on the weight parameters of the local models of the users on the ground to determine the number of quantization bits corresponding to the weight parameters of each local model specifically includes: For each ground user, the weight parameter when the local model of the ground user is updated is expressed by the following formula: ; A quantized model analysis is performed on the local model of the ground user, and the quantized sequence of the weight parameters of the local model of the ground user is: ; Determine the maximum and minimum values ​​from the quantized sequence and determine the quantized sequence interval: ; The number of quantized bits corresponding to the ground user For variables, determine the quantization interval representation: ; The quantized sequence interval is quantized according to the quantized interval representation: , ,… ,… , , ; When the weight parameters of the local model of the ground user fall into the quantization representation of the quantization sequence interval, the quantization function is: , , , ; According to the obtained quantization function, the number of quantization bits corresponding to the weight parameter of the local model of the ground user is expressed as: ; in, is the weight parameter of the local model of the nth ground user, is the original weight parameter before updating, is the learning rate for updating the weight parameters, is the loss function after gradient calculation, N is the total number of ground users, For the communication task of the nth ground user, is the value of the dth dimension of the quantized sequence of the weight parameters of the local model of the nth ground user, is the minimum value in the quantized sequence of the weight parameters of the local model of the nth ground user, is the maximum value in the quantized sequence of the weight parameters of the local model of the nth ground user, is the number of quantization bits corresponding to the nth terrestrial user, is the quantization interval, is the quantization function of the weight parameters of the local model of the nth ground user, represents the symbolic function, According to probability, is the number of quantization bits corresponding to the weight parameter of the local model of the nth ground user, Represents the number of dimensions of the neural network.

4. The satellite-ground fusion network scheduling method according to claim 1, characterized in that: The step of determining the information transmission rate representation of the ground user unloading local quantization model parameters according to the transmission link graph representation between the ground user and the satellite specifically includes: The information transmission rate expression of the ground user offloading local quantization model parameters is determined by the following formula based on the transmission link diagram between the ground user and the satellite: ； in, is the transmission power of the ground user, is the transmission bandwidth from ground users to satellites, is the noise power, It is terrestrial users and The channel gain between satellites, It is terrestrial users and Channel interference from other regions during data transmission between satellites, For the The information transmission rate of the local quantitative model parameters offloaded by each ground user.

5. The satellite-ground fusion network scheduling method according to claim 1, characterized in that: The method of determining the total delay of the federated graph learning for satellite-ground fusion network scheduling according to the local computing time representation and transmission time representation of each ground user performing the communication task specifically includes: The total delay of federated graph learning for satellite-ground fusion network scheduling is determined by the following formula based on the local computing time representation and transmission time representation of each ground user performing the communication task: ; in, is the total delay of federated graph learning, The local computation time to execute the task for the nth ground user, For the communication task of the nth ground user, It is the information transmission rate between ground users and satellites.

6. The satellite-ground fusion network scheduling method according to claim 1, characterized in that: The method of learning to transfer and aggregate node information in a dynamic satellite-ground fusion network through spatial graphs to establish a transmission link graph, and solving a federated graph optimization model according to the established transmission link graph through deep federated reinforcement learning, specifically includes: Through spatial graph learning, the node information in the satellite-ground fusion network is transferred and aggregated based on the following formula, and the graph type between ground users and satellites in the satellite-ground fusion network is obtained: ， ； According to the graph type between ground users and satellites, deep federated reinforcement learning is used to solve the federated graph optimization model by taking the weight parameters of local models of ground users and global weight parameters of federated graph learning as states and optimizing variables as actions. The reward is determined according to the total delay of federated graph learning. The convolution operation of the spatial graph learning neural network is: ; in, The graph type between ground users and satellites in the satellite-ground fusion network. learning neural network activation functions for spatial graphs, is the node matrix, Learn the weights of the kth layer of a neural network for a spatial graph, Bias for the kth layer of the spatial graph learning neural network, is the degree matrix of each node, is the adjacency feature matrix, is the convolution operation, is a graph node, is the convolution kernel, is the image pixel matrix, is the convolution kernel matrix.

7. The satellite-ground fusion network scheduling method according to claim 6, characterized in that: The method uses deep federated reinforcement learning, takes the weight parameters of local models of users on each floor and the global weight parameters of federated graph learning as states, takes optimized variables as actions, determines rewards according to the total delay of federated graph learning, and solves the federated graph optimization model, specifically including: The Actor-Critic network is used to solve the optimization variables. The status is: ; The Actor-Critic network performs each iteration The actions are: ; The Actor-Critic network performs each iteration The rewards are: ; in, Learn global weight parameters for the federated graph, is the weight parameter of the local model of each user on the ground, is the CPU cycle frequency of the nth ground user, is the transmission bandwidth between the nth ground user and the satellite, is the number of quantization bits used to quantize the local model of the nth terrestrial user, Total latency for learning the federated graph.

8. A satellite-ground fusion network scheduling device, characterized in that: include: A satellite-ground fusion network of multiple ground users, a central server and multiple satellites; the ground users use satellites as unloading units to unload local model parameters to the central server through the satellites to perform federated graph learning for satellite-ground fusion network scheduling; Each ground user builds a local model for task processing and executes it locally based on the communication task and the CPU cycle frequency of each ground user as the optimization target; A scheduling module, used to determine the local computing time representation of each ground user to perform the communication task; The weight parameters of the local model of each ground user are quantitatively analyzed to determine the number of quantization bits corresponding to the weight parameters of each local model to determine the quantized local quantization model parameters; according to the transmission link diagram between the ground user and the satellite, the information transmission rate representation of the ground user unloading the local quantization model parameters is determined; determining a transmission time representation for each ground user based on the information transmission rate representation for each ground user; The total delay of federated graph learning for satellite-ground fusion network scheduling is determined based on the local computing time representation and transmission time representation of each ground user performing communication tasks. The CPU cycle frequency of each ground user, the transmission bandwidth between each ground user and the satellite, the number of quantized bits of each ground user, and the transmission link graph between the ground user and the satellite are used as optimization variables, and the total delay of federated graph learning is minimized as the optimization goal to construct a federated graph optimization model. The node information in the dynamic satellite-ground fusion network is transferred and aggregated through spatial graph learning to establish a transmission link graph. The federated graph optimization model is solved according to the established transmission link graph through deep federated reinforcement learning, and each ground user and each satellite performs federated graph learning under the working parameters corresponding to the solved optimization variables to perform network scheduling.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

Images