A method and system for constructing and encoding scheduling computing networks for satellite-to-ground semantic communication.

Abstract

This invention relates to a computing power network construction and coding scheduling method for satellite-to-ground semantic communication, comprising the following steps: S1: Constructing a computing power network composed of LEO satellites, cloud servers, edge devices, and terminal devices; uniformly abstracting and modeling the heterogeneous computing, communication, and storage resources of nodes in the computing power network; S2: Encoding the adaptive joint source channel with dual-awareness of channel and computing power at the node level of the computing power network; realizing dynamic adaptive adjustment of the node-level semantic transmission rate according to channel conditions and computing power status; S3: Scheduling system-level resources of the computing power network to realize distributed, adaptive scheduling and resource coordination of system-level semantic tasks in the computing power network. This invention proposes adaptive joint source channel coding with dual-awareness of channel and computing power, realizing dynamic adaptive adjustment of semantic feature transmission volume by introducing a dual-awareness mechanism of channel state and device computing power.

Description

Technical Field

[0001] This invention relates to the field of semantic communication technology, specifically to a computing network construction and coding scheduling system for satellite-to-ground semantic communication. Background Technology

[0002] The sixth-generation (6G) mobile communication network aims to achieve ubiquitous coverage, connectivity, ultra-low latency, and ultra-high reliability. Within this grand vision, Low Earth Orbit (LEO) satellites, with their advantages of wide coverage, low propagation latency, and flexible networking, have become the core pillar for building Satellite-Terrestrial Integrated Networks (STINs). STINs are crucial for filling communication gaps in over 80% of the world's land area and 95% of its ocean area, and are a key technology for realizing truly ubiquitous global connectivity with 6G.

[0003] Meanwhile, semantic communication, as a revolutionary communication paradigm, is receiving widespread attention from academia and industry. Unlike traditional communication, which only ensures the accurate transmission of bit streams, semantic communication focuses on directly transmitting semantic information related to the task objective, thereby fundamentally improving bandwidth utilization efficiency and exhibiting excellent robustness under harsh channel conditions such as low signal-to-noise ratio. This task-oriented characteristic is highly compatible with the requirements of 6G for efficient and intelligent information transmission. Therefore, integrating semantic communication into STINs will open up broad prospects for achieving high-performance and intelligent applications globally.

[0004] Despite its immense potential, the integration of semantic communication with STINs still faces significant technical challenges, primarily in two aspects. First, the satellite-to-ground channel environment is highly dynamic and unstable. Doppler shift, shadowing effects, and compound fading severely impact channel quality, causing drastic fluctuations in channel conditions, resulting in significant differences from the terrestrial wireless environment. Second, the computing resources of satellite-to-ground network nodes are limited. End-to-end semantic communication heavily relies on complex deep learning models for real-time semantic encoding and decoding. These models typically have enormous computational and parameter counts, requiring substantial computing resources. However, LEO satellites are limited by size, weight, and power consumption; ground terminal equipment faces similar challenges.

[0005] Existing technologies typically rely solely on channel conditions to adjust transmission rates. Given the limited computing power of both satellites and terminal devices, this single channel-aware mechanism is insufficient to address real-world challenges in STINs. A deeper problem lies in the lack of a unified architecture to perceive, abstract, and manage heterogeneous computing resources between satellite and ground stations, leading to a disconnect between communication and computation. When tasks run under high load, even with favorable channel conditions, insufficient node computing power or uneven system-level task allocation can significantly increase inference latency, thereby compromising end-to-end quality of service. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a satellite-ground integrated semantic communication method and system based on computing power networks. This solution integrates innovations in system architecture, node technology, and scheduling strategies.

[0007] To achieve the above objectives, the present invention provides a satellite-to-ground fusion semantic communication method based on computing power networks, comprising the following steps:

[0008] S1: Construct a computing power network consisting of LEO satellites, cloud servers, edge devices, and terminal devices; perform unified abstraction and modeling of the heterogeneous computing, communication, and storage resources of nodes in the computing power network to form a globally perceptible and schedulable virtualized computing power resource pool, providing a unified interface for access to computing power services and resource allocation for upper-layer semantic communication tasks;

[0009] S2: Encode the adaptive joint source channel of the computing power network node-level channel with dual perception of computing power; realize the dynamic adaptive adjustment of the node-level semantic transmission rate according to channel conditions and computing power status;

[0010] S3: Schedule computing network system-level resources to achieve distributed, adaptive scheduling and resource collaboration of system-level semantic tasks in the computing network.

[0011] Furthermore, step S2 specifically includes the following steps:

[0012] S2.1: The transmitting end monitors the channel status information of its communication link and the computing load information of this device in real time;

[0013] S2.2: The channel state information and the computational load information are input together into a feature selection module, which dynamically determines the number of semantic features to be transmitted;

[0014] S2.3: Based on the dynamic decision, select the semantic features extracted from the original data, and process and transmit them using a joint source-channel coding method;

[0015] S2.4: The receiving end performs joint source-channel decoding on the received signal to reconstruct the semantic representation or task result that meets the requirements of the target semantic task;

[0016] The computational load information includes, but is not limited to, GPU utilization, processor utilization, task queue length, or inference latency; the feature selection module is composed of a multi-layer perceptron.

[0017] Furthermore, step S3 specifically includes the following steps:

[0018] S3.1: The terminals, LEO satellites, edge nodes, and cloud servers in the computing power network are all modeled as independent intelligent agents;

[0019] S3.2: Each agent independently makes a decision on the execution location of the semantic task, i.e., a task unloading decision, based on its locally observed state information;

[0020] S3.3: A multi-agent proximal policy optimization algorithm is adopted to collaboratively train and optimize the policies of all the agents. The training objective is to maximize a global reward function that comprehensively considers task latency, transmission quality and system load balancing.

[0021] The status information includes task latency requirements, channel status, the node's own real-time computing load, and task queue length.

[0022] Furthermore, the multi-agent proximal policy optimization algorithm adopts a framework of centralized training and distributed execution, in which the central trainer collects the experience data of all agents to collaboratively update the network parameters, and each agent makes independent decisions based on local observations, thereby achieving decentralized collaborative optimization.

[0023] Furthermore, the global reward function is:

[0024]

[0025] in Indicates the end-to-end latency of the task. It refers to the quality of semantic task completion or semantic reconstruction. It is the system load balancing degree. , , These are weighting coefficients used to balance different optimization objectives.

[0026] This invention also relates to a computing network construction and coding scheduling system for satellite-to-ground semantic communication, comprising a computer module that applies the aforementioned computing network construction and coding scheduling method for satellite-to-ground semantic communication.

[0027] Furthermore, the aforementioned space-ground fusion semantic communication system based on computing power networks includes:

[0028] The computing power network resource pool construction module performs unified abstraction and modeling of heterogeneous computing, communication and storage resources of LEO satellites, cloud servers, edge devices and terminal devices to form a globally perceptible and schedulable virtualized computing power resource pool;

[0029] The node-level channel computing power dual-aware adaptive joint source channel coding module, based on channel state information and node computing load information, dynamically decides the semantic features to be transmitted through the feature selection module, and performs joint source channel coding and decoding to achieve adaptive adjustment of the node-level semantic transmission rate according to channel conditions and computing power status.

[0030] The system-level resource scheduling module models terminals, LEO satellites, edge nodes, and cloud servers as independent intelligent agents. It uses a multi-agent near-end strategy optimization algorithm to schedule the execution location of semantic tasks, thereby realizing distributed, adaptive scheduling and resource collaboration of system-level semantic tasks in the computing network.

[0031] Furthermore, the joint source-channel coding includes a semantic source encoder and a channel encoder at the transmitting end, and a channel decoder and a semantic source decoder at the receiving end. The semantic source encoder is implemented using a deep convolutional network, which extracts multi-dimensional semantic feature maps from the original input image through convolution, batch normalization, ReLU activation, and downsampling operations to achieve efficient extraction of semantic features. Both the channel encoder and the channel decoder adopt a ResNet-based structure. The semantic source decoder gradually reconstructs the original data through a network structure that includes convolutional layers, ResNet, deconvolutional layers, layer normalization, ReLU activation function, and Sigmoid activation function.

[0032] Furthermore, the node-level channel computing power dual-sensing adaptive joint source channel coding module adopts an end-to-end training method, and jointly optimizes the image reconstruction quality, the number of transmitted features, and the computational cost through a multi-objective loss function, wherein the multi-objective loss function includes reconstruction loss, feature number penalty term, and computational cost constraint term.

[0033] Furthermore, the end-to-end training method is carried out in a simulated satellite-to-ground channel environment, with the signal-to-noise ratio (SNR) varying uniformly within the range of 0-20dB and the GPU utilization varying within the range of 0%-95%, in order to adapt to the dynamic satellite-to-ground environment.

[0034] Beneficial effects

[0035] This invention proposes a computing network construction and coding scheduling method and system for STINs (Spatial-to-Ground Semantic Communication). This invention constructs an integrated computing network encompassing satellite, cloud, edge, and terminal, providing a unified resource management and scheduling foundation for upper-layer semantic communication tasks.

[0036] Based on a computing power network architecture, this invention proposes innovative methods at two levels: node transmission and system scheduling. At the node transmission level, this invention proposes adaptive joint source-channel coding with dual-awareness of channel state and device computing power. By introducing a dual-awareness mechanism of channel state and device computing power, dynamic adaptive adjustment of semantic feature transmission volume is achieved. This method enables each node to ensure reliability by increasing the number of features under adverse channel conditions, and to prioritize low latency by reducing the number of features under high computational load, thus achieving an intelligent balance between transmission quality and processing efficiency in complex satellite-to-ground dynamic environments. At the system scheduling level, this invention designs a distributed resource scheduling strategy based on MAPPO. This strategy models each node in the network as an intelligent agent, achieving optimal allocation of semantic tasks in the computing power network through decentralized collaborative learning. This not only avoids the bottleneck of a single control node but also enables the system to adapt to the drastic fluctuations of satellite-to-ground links and the dynamic changes in computational load, achieving self-organization and self-optimization of global resources. Simulation results show that this invention performs excellently in CIFAR-10 image transmission tasks under simulated real satellite-to-ground channel conditions. Attached Figure Description

[0037] Figure 1 This is a framework diagram of the computing network construction and coding scheduling system for satellite-to-ground semantic communication according to the present invention.

[0038] Figure 2 This is an architecture diagram of the adaptive joint source-channel coding method (StarJSCC) based on dual-sensory channel computing power of the present invention.

[0039] Figure 3 This is a framework diagram of the distributed resource scheduling strategy based on MAPPO of the present invention.

[0040] Figure 4 The simulation diagram shows the PSNR performance comparison of different satellite-to-ground channel methods of the present invention under various signal-to-noise ratios.

[0041] Figure 5 The simulation diagram shows the comparison of inference latency performance of different satellite-to-ground channel methods of the present invention under various GPU utilization rates. Detailed Implementation

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0043] The present invention provides a satellite-to-ground fusion semantic communication method based on computing power networks, comprising the following steps:

[0044] S1: Construct a computing power network consisting of LEO satellites, cloud servers, edge devices, and terminal devices; perform unified abstraction and modeling of the heterogeneous computing, communication, and storage resources of nodes in the computing power network to form a globally perceptible and schedulable virtualized computing power resource pool, providing a unified interface for access to computing power services and resource allocation for upper-layer semantic communication tasks;

[0045] S2: Encode the adaptive joint source channel of the computing power network node-level channel with dual perception of computing power; realize the dynamic adaptive adjustment of the node-level semantic transmission rate according to channel conditions and computing power status. Specifically, it includes the following steps:

[0046] S2.1: The transmitting end monitors the channel status information of its communication link and the computing load information of this device in real time;

[0047] S2.2: The channel state information and the computational load information are input together into a feature selection module, which dynamically determines the number of semantic features to be transmitted;

[0048] S2.3: Based on the dynamic decision, select the semantic features extracted from the original data, and process and transmit them using a joint source-channel coding method;

[0049] S2.4: The receiving end performs joint source-channel decoding on the received signal to reconstruct the semantic representation or task result that meets the requirements of the target semantic task;

[0050] The computational load information includes, but is not limited to, GPU utilization, processor utilization, task queue length, or inference latency; the feature selection module is composed of a multi-layer perceptron.

[0051] S3: Schedule computing network system-level resources to achieve distributed, adaptive scheduling and resource coordination of system-level semantic tasks within the computing network. Specifically, this includes the following steps:

[0052] S3.1: The terminals, LEO satellites, edge nodes, and cloud servers in the computing power network are all modeled as independent intelligent agents;

[0053] S3.2: Each agent independently makes a decision on the execution location of the semantic task, i.e., a task unloading decision, based on its locally observed state information;

[0054] S3.3: A multi-agent proximal policy optimization algorithm is adopted to collaboratively train and optimize the policies of all the agents. The training objective is to maximize a global reward function that comprehensively considers task latency, transmission quality and system load balancing.

[0055] The status information includes task latency requirements, channel status, the node's own real-time computing load, and task queue length. The multi-agent proximal policy optimization algorithm adopts a framework of centralized training and distributed execution. The central trainer collects the experience data of all agents to collaboratively update the network parameters, and each agent makes independent decisions based on local observations, achieving decentralized collaborative optimization.

[0056] The global reward function is:

[0057]

[0058] in, Indicates the end-to-end latency of the task. It refers to the quality of semantic task completion or semantic reconstruction. It is the system load balancing degree. , , These are weighting coefficients used to balance different optimization objectives.

[0059] This invention also relates to a computing network construction and coding scheduling system for satellite-to-ground semantic communication, comprising a computer module that applies the aforementioned computing network construction and coding scheduling method for satellite-to-ground semantic communication.

[0060] This invention also relates to a satellite-ground fusion semantic communication system based on computing power networks, comprising:

[0061] The computing power network resource pool construction module performs unified abstraction and modeling of heterogeneous computing, communication and storage resources of LEO satellites, cloud servers, edge devices and terminal devices to form a globally perceptible and schedulable virtualized computing power resource pool;

[0062] A node-level adaptive joint source-channel coding module with dual-awareness of channel computing power dynamically determines the semantic features to be transmitted based on channel state information and node computing load information through a feature selection module, and performs joint source-channel coding and decoding to achieve adaptive adjustment of the node-level semantic transmission rate according to channel conditions and computing power status. The joint source-channel coding includes a semantic source encoder and a channel encoder at the transmitting end, and a channel decoder and a semantic source decoder at the receiving end. The semantic source encoder is implemented using a deep convolutional network, extracting multi-dimensional semantic feature maps from the original input image through convolution, batch normalization, ReLU activation, and downsampling operations to achieve efficient semantic feature extraction. Both the channel encoder and channel decoder adopt a ResNet-based structure. The semantic source decoder gradually reconstructs the original data through a network structure containing convolutional layers, ResNet, deconvolutional layers, layer normalization, ReLU activation function, and Sigmoid activation function.

[0063] The node-level channel computing power dual-sensing adaptive joint source-channel coding module adopts an end-to-end training method. It jointly optimizes image reconstruction quality, number of transmitted features, and computational cost through a multi-objective loss function, which includes reconstruction loss, feature quantity penalty term, and computational cost constraint term.

[0064] The end-to-end training method is carried out in a simulated satellite-to-ground channel environment, with the signal-to-noise ratio (SNR) varying uniformly within the range of 0-20dB and the GPU utilization varying within the range of 0%-95%, in order to adapt to the dynamic satellite-to-ground environment.

[0065] The system-level resource scheduling module models terminals, LEO satellites, edge nodes, and cloud servers as independent intelligent agents. It uses a multi-agent near-end strategy optimization algorithm to schedule the execution location of semantic tasks, thereby realizing distributed, adaptive scheduling and resource collaboration of system-level semantic tasks in the computing network.

[0066] Example 1

[0067] like Figure 1 As shown, this invention constructs an integrated computing power network encompassing satellite, cloud, edge, and terminal. This network integrates heterogeneous resources from LEO satellites, cloud servers, edge devices, and terminal devices. The heterogeneous computing, communication, and storage resources of these nodes are uniformly abstracted and modeled to form a globally perceptible and schedulable virtualized computing power resource pool, providing a unified interface for accessing computing power services and allocating resources for upper-layer semantic communication tasks. Under this architecture, the division of labor among the layers is as follows:

[0068] Satellite nodes are the core relays in the space-to-ground link, undertaking global coverage and semantic data forwarding tasks. Due to limitations in onboard computing power, satellites are suitable for performing lightweight semantic processing operations, such as feature compression and partial semantic recovery, to ensure real-time performance in highly dynamic links.

[0069] As a global computing center, the ground-based cloud server, with its ample computing power, can perform highly complex processes such as deep semantic reasoning, cross-task fusion, and model optimization, and provides global scheduling and optimization decision support for the entire computing network.

[0070] Edge nodes are located between the terminal and the cloud, and can take advantage of both in terms of geographical location and resource conditions. They are suitable for undertaking medium-complexity tasks such as encoding, decoding, and semantic compression, thereby reducing the latency of cross-regional transmission and alleviating the load on the cloud.

[0071] Terminal devices, as the source and destination of semantic information, are responsible for data collection, preliminary semantic encoding and feature extraction, and decide whether to offload the task to satellite, edge or cloud based on their local computing power and channel conditions.

[0072] Example 2

[0073] The complete architecture of dual-aware adaptive joint source-channel coding (StarJSCC) is as follows: Figure 2 As shown, this includes the collaborative workflow of the sending and receiving ends.

[0074] 2.1 Sending end processing flow

[0075] 2.1.1 Semantic Feature Extraction

[0076] The semantic source encoder at the sending end is implemented using a deep convolutional network. This network extracts multi-dimensional feature maps containing rich semantic information from the original input image through convolution, batch normalization, ReLU activation, and downsampling operations.

[0077] 2.1.2 Dual-sensory feature selection

[0078] The system monitors two key metrics in real time: the signal-to-noise ratio (SNR) of the current communication link and the current GPU utilization of the device. The normalized [SNR, GPU utilization] vector is input into a feature selection module implemented by a multilayer perceptron. This feature selection module generates a feature selection mask based on the input communication and computing status, and the features to be transmitted can be selected based on this mask.

[0079] This dual-sensory feature selection mechanism can intelligently balance communication and computational overhead: under high GPU load, it significantly reduces computational overhead and ensures low latency by selectively transmitting a small number of the most critical features; under low SNR channel conditions, it increases information redundancy by transmitting a larger number of features, thereby ensuring robust transmission and reconstruction quality.

[0080] 2.1.3 Joint source-channel coding

[0081] The selected features are fed into a ResNet-based channel encoder, and the channel-coded signal is then transmitted through a simulated satellite-to-ground channel. The simulated satellite-to-ground channel in this invention uses the Corazza model and incorporates the Doppler frequency shift effect to match the dynamic characteristics of the satellite-to-ground link.

[0082] 2.2 Receiver Processing Flow

[0083] 2.2.1 Channel Decoding and Feature Reconstruction

[0084] The receiver first performs channel decoding on the received signal. The channel decoder uses a ResNet-based architecture and aims to recover compressed semantic features from the damaged signal.

[0085] 2.2.2 Semantic Feature Decoding and Image Reconstruction

[0086] The reconstructed semantic features are input into the semantic source decoder. This decoder reconstructs the original image step by step through convolutional layers, ResNet, deconvolutional layers, layer normalization, ReLU activation function, and finally convolutional layers and sigmoid activation function.

[0087] 2.2.3 End-to-end training mechanism

[0088] The entire StarJSCC system (including the semantic encoder, feature selection module, channel encoder, channel decoder, and semantic decoder) is trained end-to-end. The training objective is to jointly optimize image reconstruction quality, the number of transmitted features, and computational cost through a multi-objective loss function.

[0089] Example 3

[0090] To efficiently schedule resources in a computing network, this invention designs a distributed resource scheduling strategy based on MAPPO, the architecture of which is as follows: Figure 3 As shown.

[0091] The terminals, LEO satellites, edge nodes, and cloud servers in the computing network are all modeled as independent intelligent agents, collectively forming a multi-agent system that collaborates to achieve a global optimization goal in a shared environment. Each agent's local observation state includes: task latency requirements, channel state, the node's own real-time computing load, and task queue length. Each agent's action is a discrete decision, representing the current execution position of the semantic task.

[0092] The system employs a "centralized training, distributed execution" framework. During the training phase, the central trainer collects experience data from all agents and collaboratively updates the network parameters; during the execution phase, each agent makes independent decisions based on its local observations.

[0093] Effect verification

[0094] System training and performance simulation analysis

[0095] End-to-end training setup

[0096] End-to-end joint training of StarJSCC was performed: the training environment simulated a uniform SNR variation within the range of 0-20dB; the simulated GPU utilization range was set to 0%-95%. The adaptive joint source-channel coding with dual-sensory channel computing power (StarJSCC) adopts an end-to-end multi-objective loss function, which integrates three objectives: maximizing image reconstruction quality, minimizing the number of transmitted features, and constraining computational overhead through a penalty term.

[0097] Simulation parameter configuration

[0098] To verify the system performance, simulations were performed on the CIFAR-10 dataset, with the key parameter settings as follows:

[0099] Table 1 Simulation parameters of satellite-to-ground channel

[0100] Parameter symbol Parameter value Parameter Description 20 GHz (Ka band) carrier frequency 1 µs Sampling interval 0 – 20 dB Signal-to-noise ratio range 10 Rice K factor 0 dB Mean of shadow fading 3 dB Shadow fading standard deviation ±500 kHz Doppler shift

[0101] Performance Analysis

[0102] Simulation results are as follows Figure 4 and Figure 5 As shown, it includes the following:

[0103] (1) Reconstruction quality: such as Figure 4 As shown, under different SNRs, the image reconstruction quality (PSNR) of the proposed dual-sensory adaptive joint source-channel coding (StarJSCC) remains robust. At low SNR (0dB), its PSNR is approximately 22.5dB, significantly better than the traditional BPG+LDPC scheme and comparable to the performance of SwinJSCC. At high SNR (20dB), its PSNR reaches 28.44dB, slightly lower than BPG+LDPC, but significantly better than other semantic communication baselines.

[0104] (2) Reasoning delay: such as Figure 5 As shown, under different GPU utilization rates, the proposed dual-sensory adaptive joint source-channel coding (StarJSCC) consistently exhibits the lowest inference latency. At a high GPU load of 95%, the total latency for processing 10,000 test images is only 5.64 seconds, approximately 20% lower than that of SwinJSCC.

[0105] The above results demonstrate that by constructing a computing power network and implementing node-level adaptive joint source channel coding (StarJSCC) with dual perception of channel computing power and system-level resource scheduling, this invention effectively solves the challenges posed by dynamic channels and limited computing power in space-ground converged networks, providing a practical solution for 6G global intelligent connectivity.

[0106] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for constructing and encoding scheduling a computing network for satellite-to-ground semantic communication, characterized in that, Includes the following steps: S1: Construct a computing power network consisting of LEO satellites, cloud servers, edge devices, and terminal devices; perform unified abstraction and modeling of the heterogeneous computing, communication, and storage resources of nodes in the computing power network, and collect real-time computing load and communication status information of each node to form a globally perceptible and schedulable virtualized computing power resource pool, providing a unified interface for computing power service access and resource allocation for upper-layer semantic communication tasks. S2: Encode the adaptive joint source channel of the computing power network node-level channel with dual perception of computing power; realize the dynamic adaptive adjustment of the node-level semantic transmission rate according to channel conditions and computing power status; S2.1: The transmitting end monitors the channel status information of its communication link and the computing load information of this device in real time; S2.2: The channel state information and the computational load information are input together into a feature selection module, which dynamically determines the number of semantic features to be transmitted; S2.3: Based on the dynamic decision, select the semantic features extracted from the original data, and process and transmit them using a joint source-channel coding method; S2.4: The receiving end performs joint source-channel decoding on the received signal to reconstruct the semantic representation or task result that meets the requirements of the target semantic task; The computational load information includes GPU utilization, processor utilization, task queue length, or inference latency; the feature selection module is composed of a multi-layer perceptron. S3: Schedule computing network system-level resources to achieve distributed, adaptive scheduling and resource collaboration of system-level semantic tasks in the computing network.

2. The computing network construction and coding scheduling method for satellite-to-ground semantic communication according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1: The terminals, LEO satellites, edge nodes, and cloud servers in the computing power network are all modeled as independent intelligent agents; S3.2: Each agent independently makes a decision on the execution location of the semantic task, i.e., a task unloading decision, based on its locally observed state information; S3.3: A multi-agent proximal policy optimization algorithm is adopted to collaboratively train and optimize the policies of all the agents. The training objective is to maximize a global reward function that comprehensively considers task latency, transmission quality and system load balancing. The status information includes task latency requirements, channel status, the node's own real-time computing load, and task queue length.

3. The computing network construction and coding scheduling method for satellite-to-ground semantic communication according to claim 2, characterized in that, The multi-agent proximal policy optimization algorithm adopts a framework of centralized training and distributed execution. The central trainer collects the experience data of all agents to collaboratively update the network parameters, and each agent makes independent decisions based on local observations, thereby achieving decentralized collaborative optimization.

4. The computing network construction and coding scheduling method for satellite-to-ground semantic communication according to claim 2, characterized in that, The global reward function is: in Indicates the end-to-end latency of the task. It refers to the quality of semantic task completion or semantic reconstruction. It is the system load balancing degree. , , These are weighting coefficients used to balance different optimization objectives.

5. A satellite-to-ground semantic communication system based on a computing power network, characterized in that, It includes a computer module that applies the satellite-to-ground semantic communication method based on computing power network according to any one of claims 1 to 4.

6. The computing network construction and coding scheduling system for satellite-to-ground semantic communication according to claim 5, characterized in that, include: The computing power network resource pool construction module performs unified abstraction and modeling of heterogeneous computing, communication and storage resources of LEO satellites, cloud servers, edge devices and terminal devices to form a globally perceptible and schedulable virtualized computing power resource pool; The node-level channel computing power dual-aware adaptive joint source channel coding module, based on channel state information and node computing load information, dynamically decides the semantic features to be transmitted through the feature selection module, and performs joint source channel coding and decoding to achieve adaptive adjustment of the node-level semantic transmission rate according to channel conditions and computing power status. The system-level resource scheduling module models terminals, LEO satellites, edge nodes, and cloud servers as independent intelligent agents. It uses a multi-agent near-end strategy optimization algorithm to schedule the execution location of semantic tasks, thereby realizing distributed, adaptive scheduling and resource collaboration of system-level semantic tasks in the computing network.

7. The computing network construction and coding scheduling system for satellite-to-ground semantic communication according to claim 6, characterized in that, The joint source-channel coding includes a semantic source encoder and a channel encoder at the transmitting end, and a channel decoder and a semantic source decoder at the receiving end. The semantic source encoder is implemented using a deep convolutional network, which extracts multi-dimensional semantic feature maps from the original input image through convolution, batch normalization, ReLU activation, and downsampling operations to achieve efficient extraction of semantic features. Both the channel encoder and the channel decoder adopt a ResNet-based structure. The semantic source decoder reconstructs the original data step by step through a network structure that includes convolutional layers, ResNet, deconvolutional layers, layer normalization, ReLU activation function, and Sigmoid activation function.

8. The computing network construction and coding scheduling system for satellite-to-ground semantic communication according to claim 6, characterized in that, The node-level channel computing power dual-sensing adaptive joint source channel coding module adopts an end-to-end training method. It jointly optimizes image reconstruction quality, number of transmitted features, and computational overhead through a multi-objective loss function, which includes reconstruction loss, feature quantity penalty term, and computational overhead constraint term.

9. The computing network construction and coding scheduling system for satellite-to-ground semantic communication according to claim 8, characterized in that, The end-to-end training method is carried out in a simulated satellite-to-ground channel environment, with the signal-to-noise ratio (SNR) varying uniformly within the range of 0-20dB and the GPU utilization varying within the range of 0%-95%, in order to adapt to the dynamic satellite-to-ground environment.

Images