Edge computing-based satellite-ground collaborative processing methods and systems

Abstract

A satellite-ground collaborative processing method and system based on edge computing is proposed. This method constructs an integrated satellite-ground cloud-edge-device collaborative architecture comprising a satellite terminal layer, a ground edge layer, and a cloud computing layer. Distributed ground-based edge service stations connect the satellite terminal and the ground cloud center, enabling unified scheduling of satellite and ground resources and task collaboration. A task processing location decision model is built based on task data volume, link bandwidth, and computing power. Combining response time constraints and energy consumption costs, the model dynamically selects the satellite terminal, edge station, or cloud center for processing, optimizing latency and energy consumption. A geoscientific-space intelligent learning framework is constructed. The ground center offline trains and optimizes the deep learning model, which is then pruned and quantized to generate a lightweight model that is sent to the satellite terminal. The satellite terminal performs online inference and collects data for feedback, forming a closed-loop learning process. An offline-online collaborative mode is introduced, prioritizing online processing of high real-time tasks and batch-downloading non-real-time tasks, dynamically adjusting strategies to improve system efficiency and resource utilization.

Description

Technical Field

[0001] This invention relates to satellite internet, edge computing and distributed intelligent systems, and in particular to a satellite-ground collaborative processing method and system based on edge computing. Background Technology

[0002] Existing satellite-ground collaborative processing architectures primarily rely on centralized ground-based processing. Satellite nodes typically only handle data acquisition and simple preprocessing, with the vast majority of data and computational tasks requiring downlink to the ground for centralized processing. While this architecture facilitates unified management and resource allocation, it struggles to fully leverage the on-orbit processing capabilities of satellites and is poorly suited for complex scenarios involving high timeliness, low latency, and massive data processing. Furthermore, current edge computing solutions mainly focus on terrestrial networks, with research on on-orbit satellite edge computing and satellite-ground resource collaboration still in its early stages.

[0003] Existing satellite-ground collaborative processing architectures typically employ a layered structure, including satellites, ground receiving stations, and a central data processing center. The satellites are primarily responsible for data acquisition and initial preprocessing, subsequently transmitting the raw or pre-processed data to the ground via satellite-ground links. Ground receiving stations receive and buffer the data, which is then ultimately processed, stored, and distributed for application by the ground center. The overall process is ground-driven, with data forwarding being the primary method of communication between satellites and ground stations. Satellite nodes lack efficient edge computing and intelligent collaborative capabilities. This architecture struggles to respond to large-scale distributed tasks in real-time and offers limited support for resource scheduling in low-latency, high-bandwidth, and dynamic environments.

[0004] Existing space-ground collaborative processing architectures suffer from several technical shortcomings: First, the limited edge computing capabilities of satellite nodes hinder on-site data analysis and intelligent processing, leading to high task processing latency. Second, the need for massive data downlinks to ground centers increases bandwidth burden and communication latency. Third, the lack of effective space-ground resource coordination and dynamic scheduling mechanisms prevents full utilization of distributed computing resources, resulting in poor overall system resilience and robustness. These deficiencies limit the effectiveness of integrated space-ground systems in scenarios requiring high timeliness and handling complex tasks. Existing space-ground collaborative processing architectures generally suffer from weak satellite-side computing capabilities and low efficiency in resource coordination and task response.

[0005] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology, and to provide a satellite-ground collaborative processing method and system based on edge computing, thereby constructing an efficient, intelligent, and adaptive satellite-ground collaborative computing system to maximize task processing efficiency under limited satellite-ground resources.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect of the present invention, a satellite-ground collaborative processing method based on edge computing includes the following steps:

[0009] S1. Construct an integrated space-ground cloud-edge-device collaborative architecture, including a satellite terminal layer, a ground edge layer, and a cloud computing layer. Connect the satellite terminal and the ground cloud center through distributed ground-based edge service stations to form a three-layer collaborative processing system, realizing unified scheduling and task collaboration of space-ground resources;

[0010] S2. Based on the task data volume, link bandwidth and computing power, establish a task processing location decision model. According to the task response time constraints and energy consumption costs, dynamically select to process at the satellite end, ground edge station or cloud center to achieve joint optimization of task processing latency and energy consumption.

[0011] S3. Construct an intelligent learning framework for Earth-based and space-based applications. Conduct offline training and optimization of deep learning models at the ground center. Generate lightweight models through model pruning and quantization, and send them to the satellite for online inference and data collection, forming a closed-loop learning process of ground training - satellite inference - data feedback - model iteration.

[0012] S4. Introduce an offline-online collaborative processing mode, prioritize online processing of high real-time tasks on satellites or edge nodes, cache and batch download non-real-time tasks, and dynamically adjust task allocation strategies based on link status and node load to improve system response efficiency and resource utilization.

[0013] In a second aspect of the invention, a satellite-ground collaborative processing system based on edge computing is provided, comprising:

[0014] The satellite terminal layer includes multiple satellite devices with edge computing capabilities, used for data acquisition, preliminary processing, and online inference;

[0015] The ground edge layer includes multiple distributed ground-based edge service stations, which are used to receive satellite data, provide edge computing services, and communicate with satellite terminals and cloud centers;

[0016] The cloud computing layer, including ground data centers, is used for massive data storage, deep learning model training, and system resource coordination.

[0017] The task scheduling module is used to sense the resource status and network environment of each node in real time, and dynamically allocate tasks to the satellite, edge station or cloud center according to task characteristics and system status.

[0018] The intelligent learning module is used to train, optimize, distribute, and iteratively update the model, and supports lightweight inference and continuous learning on the satellite side.

[0019] The offline-online processing control module is used to dynamically select online processing or cached download strategy based on the real-time requirements of the task and the link status, thereby improving system resource utilization and response efficiency.

[0020] The present invention has the following beneficial effects:

[0021] This invention proposes a satellite-ground collaborative processing method and system based on edge computing, and designs a satellite-ground collaborative processing architecture based on edge computing. This architecture can not only improve the data processing capability, resource utilization efficiency and mission response time of the satellite-ground collaborative system, but also support data processing, resource collaboration and intelligent decision-making between satellite and ground nodes, thereby effectively improving the computing capability and mission response efficiency of the integrated satellite-ground system.

[0022] Specifically, this invention addresses three key technical challenges: first, it enables efficient edge computing capabilities for satellites in orbit, supporting local data processing and intelligent analysis to reduce data downlink pressure; second, it establishes a flexible and efficient resource collaboration and dynamic task distribution mechanism between satellite and ground, effectively improving the real-time response capability of distributed tasks; and third, it introduces an intelligent decision-making and task scheduling module, enabling the system to adaptively adjust task allocation and resource utilization based on dynamic environmental factors such as network status and node load, significantly enhancing the overall system's scalability, robustness, and intelligence.

[0023] The significant innovative contributions and advantages of this invention are mainly reflected in three dimensions: First, at the architectural design level, a hierarchical, multi-node collaborative satellite-ground edge computing architecture is proposed, clearly defining the functional division and collaborative relationships between satellite edge nodes, ground edge nodes, and ground center nodes, achieving efficient integration and task collaboration between on-orbit satellite and ground computing resources; Second, at the scheduling mechanism level, the architecture embeds a unified collaborative management and scheduling platform, which can perceive the resource status and network environment of each node in real time, supporting dynamic task distribution, elastic migration, and cross-node collaborative processing, significantly improving the processing efficiency and adaptability of the integrated satellite-ground system; Third, at the optimization capability level, through performance monitoring and operational feedback mechanisms, adaptive optimization of task distribution, resource utilization, and bandwidth allocation is achieved, enabling the system to continuously adjust the collaborative scheme according to the actual operating conditions, ensuring high timeliness, high reliability, and intelligent operation.

[0024] Compared with existing satellite-ground collaborative architectures that rely primarily on centralized ground processing or simple satellite-ground division of labor, this invention exhibits significant advantages: Through a systematic multi-level edge computing architecture design, it achieves efficient collaboration and elastic resource integration between satellite nodes and ground nodes; relying on a global resource status awareness and unified scheduling platform, it can dynamically adjust task distribution paths and resource allocation methods, significantly improving the efficiency of distributed task processing and the flexibility of system response; simultaneously, this invention supports intelligent task migration and collaborative processing between multiple nodes, and achieves system adaptive optimization through performance feedback, effectively reducing data downlink pressure and communication bottlenecks, enhancing the system's adaptability to complex environments and diverse business needs, and overall outperforming existing technologies in terms of scalability, robustness, and intelligence.

[0025] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0026] Figure 1 This is a flowchart of the overall process of the satellite-ground collaborative processing method based on edge computing of the present invention.

[0027] Figure 2 This is a framework diagram of the satellite orbit edge computing architecture according to an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of the star-ground-cloud-edge architecture according to an embodiment of the present invention.

[0029] Figure 4 This is the "Earth Science and Space Application" intelligent learning framework of this invention. Detailed Implementation

[0030] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0031] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0032] This invention aims to address the shortcomings of existing satellite-ground collaborative architectures, such as weak satellite-end computing power, low efficiency in resource coordination and mission response, and to improve the data processing capabilities, resource utilization efficiency, and mission response timeliness of satellite-ground collaborative systems. It proposes a three-layer collaborative architecture of "space-ground integration" (cloud-edge-device), connecting the satellite end and the ground cloud center through distributed ground-based edge service stations to achieve unified scheduling of space-ground resources and mission coordination. Simultaneously, it proposes a "geoscience-space application" intelligent learning framework and an on-orbit "offline-online" collaborative processing mode for satellites. Compared to existing technologies, this allows for dynamic adjustment of mission distribution and resource allocation, effectively reducing data downlink pressure and communication bottlenecks, and significantly enhancing system scalability, robustness, and intelligence.

[0033] See Figure 1 This invention provides a satellite-ground collaborative processing method based on edge computing, comprising the following steps:

[0034] Step S1. Construct a space-ground integrated cloud-edge-device collaborative architecture, including a satellite terminal layer, a ground edge layer, and a cloud computing layer. The satellite terminal and the ground cloud center are connected through distributed ground-based edge service stations to form a three-layer collaborative processing system, realizing unified scheduling and task collaboration of space-ground resources.

[0035] Step S2. Based on the task data volume, link bandwidth and computing power, establish a task processing location decision model. According to the task response time constraints and energy consumption costs, dynamically select whether to process at the satellite end, ground edge station or cloud center to achieve joint optimization of task processing latency and energy consumption.

[0036] In some embodiments, the task processing location decision model in step S2 includes: calculating the transmission time of task data to the ground edge station and the cloud center based on the task data volume, the bandwidth between the satellite and the edge station, and the bandwidth between the satellite and the cloud center; estimating the local computation processing time of the task at each node based on the computing capabilities of the satellite, the edge node, and the cloud center; combining the transmission time and the computation time to obtain the total completion time of the task processing at the satellite, the ground edge station, and the cloud center; comparing the total completion time of each processing path with the maximum allowable response time of the task to preliminarily determine a set of feasible processing locations; introducing a weighted cost function to calculate the comprehensive cost of each feasible processing location, wherein the weighted cost function is a linear weighted sum of latency cost and energy consumption cost; and selecting the node that minimizes the value of the weighted cost function as the final execution location of the task.

[0037] Step S3. Construct a ground-based intelligent learning framework for space applications. Conduct offline training and optimization of deep learning models at the ground center. Generate lightweight models through model pruning and quantization, and send them to the satellite for online inference and data collection, forming a closed-loop learning process of ground training - satellite inference - data feedback - model iteration.

[0038] In some embodiments, the geospatial intelligent learning framework described in step S3 includes: receiving and integrating remote sensing datasets periodically transmitted back from satellites at a ground data center, the datasets including input samples and their corresponding annotations; selecting the structure and scale of the neural network model based on the satellite's computing power and storage capacity constraints to ensure the model meets the resource limitations of satellite deployment; optimizing the model parameters using an iterative algorithm based on mini-batch stochastic gradient descent, training the model by minimizing an empirical loss function containing a regularization term; and during training, evaluating and monitoring the model performance using an independent validation dataset, and adjusting hyperparameters based on the evaluation results. The model is adjusted to obtain a trained initial model; the trained initial model is then optimized, including structured pruning based on parameter importance assessment to remove redundant connections, and low-bit-width transformation based on weight quantization to reduce model storage and computational overhead, thereby generating a lightweight model; the lightweight model is then distributed to the satellite via a satellite-to-ground link; the satellite uses the distributed model to perform online inference on the real-time collected data, and transmits the inference results and newly collected data back to the ground data center at a preset cycle; the ground data center uses the transmitted data to update the training dataset and initiates a new round of model training and optimization, forming a closed-loop iterative learning process.

[0039] In some embodiments, the structured pruning includes: identifying redundant connections in the model using an importance evaluation function, pruning them according to a preset retention ratio or threshold, retaining connections with high importance; calculating the model compression ratio after pruning, and evaluating the impact of pruning on model performance.

[0040] In some embodiments, the weight quantization includes: mapping floating-point weights to low-bit-width fixed-point integers, calculating quantization error based on the weight range and quantization bit width; and balancing model accuracy and storage overhead by adjusting the quantization bit width.

[0041] In some embodiments, the satellite-based online inference phase includes: estimating the computational complexity of inference based on the number of model layers and parameter size; predicting the inference time by combining the computing power of the onboard hardware; and collecting new data at a preset cycle and transmitting it back to the ground center for model iterative updates.

[0042] Step S4. Introduce an offline-online collaborative processing mode. Prioritize online processing of high real-time tasks on satellites or edge nodes, and cache and batch download non-real-time tasks. Combine link status and node load to dynamically adjust task allocation strategies, thereby improving system response efficiency and resource utilization.

[0043] In some embodiments, the offline-online collaborative processing mode in step S4 includes: defining a binary decision variable for each task to represent the decision state of whether the task is assigned to online processing or offline caching; constructing a scheduling optimization model with the goal of minimizing the total system processing time; setting constraints for the optimization model, including: the completion time of each online processing task meets its maximum allowable response time requirement, and the total amount of data cache occupied by all online processing tasks does not exceed the upper limit of the satellite-end cache capacity; dynamically solving the optimization model and updating the binary decision variable based on the real-time monitored satellite-to-ground link status and node load status to achieve dynamic adjustment of the task processing strategy.

[0044] In some embodiments, the objective function of the scheduling optimization model comprehensively considers the time cost required for online processing at the satellite end and the time cost required for caching the task and then downloading it to the cloud center for processing, wherein the latter includes ground cloud computing time and satellite-to-ground data transmission time; the constraints ensure that high real-time tasks receive a rapid response through online processing, while managing the use of storage resources at the satellite end through cache capacity constraints; by dynamically adjusting the binary decision variables, intelligent switching and adaptive allocation of tasks between online processing and offline cache download modes are achieved.

[0045] This invention also provides a satellite-ground collaborative processing system based on edge computing, comprising:

[0046] The satellite terminal layer includes multiple satellite devices with edge computing capabilities, used for data acquisition, preliminary processing, and online inference;

[0047] The ground edge layer includes multiple distributed ground-based edge service stations, which are used to receive satellite data, provide edge computing services, and communicate with satellite terminals and cloud centers;

[0048] The cloud computing layer, including ground data centers, is used for massive data storage, deep learning model training, and system resource coordination.

[0049] The task scheduling module is used to sense the resource status and network environment of each node in real time, and dynamically allocate tasks to the satellite, edge station or cloud center according to task characteristics and system status; the link status can be measured in real time, for example, through beacon signals, and the trend of change can be predicted by combining ephemeris data.

[0050] The intelligent learning module is used to train, optimize, distribute, and iteratively update the model, and supports lightweight inference and continuous learning on the satellite side.

[0051] The offline-online processing control module is used to dynamically select online processing or cached download strategy based on the real-time requirements of the task and the link status, thereby improving system resource utilization and response efficiency.

[0052] In some embodiments, the task scheduling module is configured to: dynamically select the task processing location based on the real-time resource status and link bandwidth of each node in the satellite, ground edge layer and cloud computing layer, and based on the task data volume, computational complexity and maximum allowable response time, through a weighted cost function, so as to realize the elastic distribution and migration of tasks between satellite and ground.

[0053] The intelligent learning module is configured to collaborate with the ground data center and satellite terminal layer to achieve centralized ground training, lightweight optimization, secure distribution between satellite and ground, online inference on the satellite, and periodic data backhaul of deep learning models, forming a closed-loop iterative learning process.

[0054] The offline-online processing control module is configured to work in conjunction with the task scheduling module and the satellite-to-ground communication link to dynamically decide on the task execution strategy based on the task's real-time requirements and the link load status. It controls high real-time tasks to be processed online on satellites or edge nodes, and caches non-real-time tasks and downloads them in batches to the cloud center for processing.

[0055] The edge computing-based satellite-ground collaborative processing method and system proposed in this invention have the following main technical advantages: Addressing the problems of existing satellite-ground collaborative architectures, which are primarily based on ground-centric centralized processing, have weak satellite-end computing capabilities, heavy data downlink burdens, and insufficient resource scheduling, the invention achieves multi-dimensional breakthroughs through systematic design: First, it innovates a "satellite-ground integrated" three-layer collaborative architecture, integrating satellite terminal layer, ground edge layer, and cloud computing layer resources with distributed ground-based edge service stations as the link, fully leveraging the complementary advantages of satellite and ground to effectively address challenges such as bandwidth limitations and latency sensitivity; Second, it constructs a globally aware elastic scheduling mechanism, establishing a decision model based on task data volume, link bandwidth, and computing power, and through... The weighted cost function achieves joint optimization of latency and energy consumption, supporting dynamic task distribution and cross-node migration; third, it proposes a "ground-based learning and space-based application" intelligent learning framework, which generates a lightweight model for deployment on satellite through centralized ground training and model pruning and quantization, and forms a closed-loop iteration with periodic data backhaul, improving the accuracy and timeliness of on-orbit inference; fourth, it innovates the satellite's on-orbit "offline-online" collaborative processing mode, prioritizing online processing of high real-time tasks on satellites or edge nodes, and batch downloading non-real-time tasks, combined with dynamic adjustment strategies for links and loads, significantly reducing data download pressure and enhancing the system's scalability, robustness, and intelligence level, which is superior to existing technologies in terms of adaptability and processing efficiency in complex scenarios.

[0056] The features, principles, and advantages of specific embodiments of the present invention are further described below.

[0057] A satellite-ground collaborative processing method and system based on edge computing is proposed. An edge computing-based satellite-ground collaborative processing architecture is designed, and the framework diagram is shown below. Figure 2 As shown.

[0058] 1. "Space-Ground Integrated" Cloud-Edge-Device Architecture Design

[0059] To address the technical bottleneck of traditional satellite-ground separation models that cannot fully leverage the synergy of space-ground collaboration, this invention innovatively designs a "space-ground integrated" cloud-edge-device collaborative architecture. This architecture overcomes the limitations of traditional separate ground data processing and satellite-based processing, which operate independently. By setting up distributed ground-based edge service stations as the key link connecting the satellite and the ground cloud center, a three-layer collaborative processing system is constructed, consisting of a cloud computing layer, an edge computing layer, and a satellite terminal layer. Figure 3 As shown, in this architecture, satellite equipment is redefined as intelligent computing nodes with dual functions. It can act as a service provider, directly pushing satellite observation imagery data and real-time processing results to users, and as a service requester, proactively leveraging the powerful computing and storage resources of distributed edge centers and cloud centers to collaboratively process large-scale data-intensive and computationally-intensive tasks. Ground-based edge service stations provide services to nearby satellites, reducing the burden of data transmission to the cloud center, while maintaining communication interconnection with other edge centers to meet the interoperability needs of all satellite equipment. The ground-based cloud computing center, as a persistent data storage center, is responsible for handling resource-intensive tasks such as high-latency-tolerant tasks and large-scale neural network parameter training, while also undertaking the management and resource coordination functions of the entire service platform. This architecture fully leverages the complementary advantages of satellite and ground resources, effectively addressing key technical challenges in high-throughput remote sensing data processing, such as limited communication bandwidth and latency sensitivity.

[0060] Based on a three-layer collaborative processing system (cloud computing layer, ground edge layer, and satellite terminal layer), a unified computing and communication model can be established to scientifically select task processing locations and evaluate system performance. Let the data size of a certain on-orbit task be D (bits), and the link bandwidth between the satellite and the ground edge station be... The link bandwidth between the satellite and the cloud center is The data transmission times are respectively The time required for the satellite, edge nodes, and cloud center to process this task on their respective computing platforms are respectively... , , Taking into account both transmission and computational overhead, the task completion time for different processing paths can be obtained: For tasks with time constraints Let its maximum allowable response time be The system can determine the task processing location based on the comparison of the three response time and latency constraints mentioned above, using the following decision rules:

[0061] like Then it is processed at the satellite end; if If the signal is not received, it is processed at the ground edge station; otherwise, it is sent to the cloud center. To comprehensively consider factors such as energy consumption, an energy consumption model can be introduced. Let the power consumption per unit time for performing the task at the satellite be... Then the satellite will perform the mission. The energy consumption is Similarly, energy consumption at the edge and in the cloud can be defined. , The dispatch center can select the processing location according to the weighted cost function: in , , These are the weighting coefficients. Selecting them makes... The smallest node is used as the execution location for tasks, achieving joint optimization of latency and energy consumption.

[0062] Further considering the system's throughput and resource utilization, the system can be analyzed over a time interval. The number of tasks completed within the time limit is recorded as follows: The throughput is defined as By introducing multi-node parallel processing, the throughput will increase approximately linearly with the number of available resources; at the same time, the above decision rules can ensure the timeliness requirements of high real-time services (such as remote sensing target detection, emergency tasks, etc.).

[0063] 2. "Earth Science and Space Application" Intelligent Learning and Task Allocation Framework

[0064] To improve the timeliness and accuracy of intelligent task processing on satellites, this invention establishes a "Ground-based Science-Space Application" intelligent learning framework, realizing the hierarchical division of labor and efficient adaptation of deep learning algorithms. For example... Figure 4 As shown, the ground data center is responsible for the offline processing of massive amounts of remote sensing data and the centralized training of deep learning models. It dynamically selects and optimizes the network structure based on satellite equipment performance and mission requirements. During the model optimization phase, techniques such as parameter pruning and model quantization are employed, taking into account the personalized data and mission characteristics of each satellite, to achieve lightweight deep learning models and ensure high inference accuracy and timeliness under limited onboard resources. The optimized model is then transmitted to the satellite via a secure and efficient link. The satellite is responsible for online inference and real-time data acquisition, and periodically transmits operational data back to achieve continuous model updates and performance improvements.

[0065] (1) Offline training phase: The ground data center receives remote sensing datasets periodically transmitted back from the satellite. ,in Indicates the first Input samples (e.g., remote sensing images, spectral data, etc.) This indicates the corresponding annotation information. To meet the deployment requirements under the satellite's hardware conditions, it is necessary to consider the computing capabilities of the satellite equipment. and storage capacity Choose a neural network model of appropriate size Ensure that the number of model parameters does not exceed the available storage. Inference overhead does not exceed computing power .

[0066] During the offline training phase, the ground center uses stochastic gradient descent to adjust the model parameters. Optimize the model to minimize its empirical risk on the training set. Let the loss function be... The regularization term is The regularization coefficient is Then the objective function is: in The number of training samples is given. During training, mini-batch stochastic gradient descent is used to iteratively update the parameters: let the current iteration number be . The learning rate is Each time, a random sample of size is drawn from the training set. small batch Calculate the gradient of this batch. And update the parameters according to the following rules:

[0067] During training, the model also needs to be validated. Typically, a portion of the data is divided into a validation set, and the model's accuracy or other evaluation metrics on the validation set are calculated. The generalization ability of the model is monitored and hyperparameters are tuned. Offline training ends when the loss function converges or the validation set performance meets expectations, and the initial model parameters are obtained. This lays the foundation for subsequent model optimization and on-orbit deployment.

[0068] (2) Model Optimization Stage: For the basic model trained offline, in order to meet the constraints of limited computing and storage resources on the satellite, it is necessary to perform structured pruning and weight quantization on the model to achieve model compression and acceleration. The pruning stage uses an importance evaluation function. To identify redundant connections, where For connection weights, This represents the gradient of the loss function with respect to the weights. A retention ratio is set based on the score of each parameter. or threshold Only the most important items are retained in the ranking. Or a score of not less than The connections that are selected are pruned, while the remaining connections are removed. Assume the model has a total of parameters before pruning. One, retain after pruning. If there are 1 parameter, then the compression ratio of the model is 1. Building upon this, weighted quantization maps the 32-bit floating-point weights to low-bit-width fixed-point integers, further reducing storage and bandwidth consumption. Let the minimum and maximum values ​​of the weights be respectively... and Quantization bit width is Then floating-point weight Mapped to fixed-point weights The formula is: The upper bound of the maximum absolute error introduced by quantization is By selecting an appropriate bit width A balance can be struck between hardware savings and inference accuracy. After dual optimization of structured pruning and weight quantization, a lightweight model is obtained that significantly reduces the parameter size while basically maintaining the inference performance of the original model, making it easy to deploy on spaceborne heterogeneous computing platforms.

[0069] (3) Inference and Data Acquisition Stage: After offline training and model optimization, the optimized lightweight model is sent to the satellite via a satellite-to-ground link. The satellite uses this model to perform online inference and periodically collects new remote sensing data. Assume the neural network has a total of... Layer, of which the first There are layers input neurons and If there are n output neurons, then the computational complexity of the reasoning process can be expressed as: That is, the sum of the multiplication and addition operations of each fully connected layer. For convolutional layers, this can be further calculated according to the kernel size. Feature map size And channel number expansion complexity estimation: in This refers to the number of multiply-accumulate operations that the onboard hardware can perform per second.

[0070] During inference, the satellite collects new data according to a preset period T, generating data batches. Let the amount of data collected in each period be... The total amount of data that needs to be transmitted back via the satellite-to-ground link per cycle is Add the inference result data. If the link bandwidth is... The time required to send back the data is (The inference results are usually small and can be ignored). After receiving new data and inference results, the ground center adds them to the training set to complete the continuous iterative update of the model. After the updated model completes offline training and optimization, it is sent back to the satellite, forming a closed-loop process of "centralized ground training - lightweight satellite inference - continuous data collection and transmission - iterative model optimization".

[0071] Through this stage of design, the system can realize real-time inference of the model on the satellite and continuous iterative optimization on the ground, which not only ensures the high real-time business requirements, but also ensures the adaptability and accuracy of the model in dynamic environments, thereby significantly improving the intelligence level and execution efficiency of on-orbit business processing.

[0072] 3. New satellite on-orbit "offline-online" processing mode

[0073] The timeliness and computational resource requirements of on-orbit processing services vary. Therefore, this invention innovatively proposes an offline-online collaborative processing mode to flexibly allocate tasks and improve overall efficiency within the satellite-ground collaborative system. For high real-time and response-sensitive services (such as emergency monitoring and target tracking), the traditional "transmit first, process later" mode is difficult to meet timeliness requirements due to satellite-ground link congestion and transmission delays. Therefore, the system prioritizes real-time inference and edge computing at the satellite end using local models or nearby ground edge nodes to complete critical applications such as event detection and target identification as quickly as possible, significantly reducing task response latency. For batch, non-real-time, or data that can be processed later, an offline caching and delayed processing strategy is adopted: satellite nodes store and preprocess data first, and then download it in batches to the ground center for in-depth analysis and archiving when the link load is low or the scheduling window is suitable, thus saving bandwidth and reducing the computational pressure on the satellite end.

[0074] To uniformly describe the scheduling decisions for offline and online tasks, let the task be... The amount of data is Real-time requirements are (i.e., the maximum allowed completion time), the processing time on the satellite is... The time required for processing and downloading from the ground to the cloud is ,in For the satellite-to-ground link bandwidth, a binary decision variable is introduced. ,when The task is processed online at the satellite or edge. The task is cached and downloaded offline. Therefore, the following optimization model can be constructed to minimize the total processing time while ensuring the buffer capacity does not exceed the limit, provided that real-time constraints are met: in This is due to the capacity limitation of the satellite-side cache. When the decision variable... At that time, the task Within the time limit Completed within the system; simultaneously, the total data volume of all online processing tasks does not exceed the satellite's available cache capacity. For The tasks are temporarily stored and then downloaded to the ground in batches at appropriate times. The completion latency of these tasks is mainly determined by the data transmission time and the ground processing time.

[0075] By combining real-time monitoring of link status and node load, the system dynamically updates decision variables during operation. This enables intelligent switching between offline and online modes. This mode fully utilizes the real-time computing power of the satellite and the centralized processing capabilities of the ground station, allowing high-priority tasks to respond quickly and low-priority tasks to be effectively aggregated and transmitted. This improves the utilization rate of onboard node resources and the efficiency of link bandwidth, while also significantly enhancing the entire satellite-ground collaborative system's ability to handle complex and ever-changing services and its potential for intelligent evolution.

[0076] In summary, this invention proposes a satellite-ground collaborative processing method and system based on edge computing. The key innovative contributions and significant advantages of this invention include:

[0077] (1) Systematic design of multi-level satellite-ground collaborative edge computing architecture

[0078] This invention proposes a hierarchical, multi-node collaborative satellite-ground edge computing architecture, which clearly defines the functional division and cooperation relationship between satellite edge nodes, ground edge nodes and ground center nodes, and realizes efficient integration and task collaboration between on-orbit satellite and ground computing resources.

[0079] (2) Global awareness and elastic scheduling mechanism for tasks and resources

[0080] The architecture incorporates a unified collaborative management and scheduling platform, which can perceive the resource status and network environment of each node in real time, and supports dynamic task distribution, elastic migration and cross-node collaborative processing, greatly improving the processing efficiency and adaptability of the space-ground integrated system.

[0081] (3) Intelligent adaptive optimization capability of satellite-ground collaborative processing

[0082] Through performance monitoring and operational feedback mechanisms, adaptive optimization of task distribution, resource utilization, and bandwidth allocation is achieved, enabling the system to continuously adjust the collaborative scheme according to the actual operating conditions, ensuring high timeliness, high reliability, and intelligent operation.

[0083] Compared to existing satellite-ground collaborative architectures that rely primarily on centralized ground processing or simple satellite-ground division of labor, this invention achieves efficient collaboration and flexible resource integration between satellite and ground nodes through a systematic multi-level edge computing architecture design. Relying on a global resource status awareness and unified scheduling platform, it can dynamically adjust task distribution paths and resource allocation methods, significantly improving the efficiency of distributed task processing and the flexibility of system response. Furthermore, this invention supports intelligent task migration and collaborative processing between multiple nodes and achieves system adaptive optimization through performance feedback, effectively reducing data downlink pressure and communication bottlenecks, enhancing the system's adaptability to complex environments and diverse business needs, and demonstrating superior scalability, robustness, and intelligence compared to existing technologies.

[0084] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A method for edge computing based on satellite-ground cooperative processing, characterized in that, The method comprises the following steps: S1. Constructing a star-ground integrated cloud edge-end collaborative architecture, including a satellite terminal layer, a ground edge layer and a cloud computing layer, connecting the satellite end and the ground cloud center through a distributed ground edge service station to form a three-layer collaborative processing system, realizing unified scheduling of star-ground resources and task collaboration; S2. Based on the task data volume, link bandwidth and computing capacity, a task processing location decision model is established, and according to the task response time constraint and energy consumption cost, the processing is dynamically selected on the satellite end, the ground edge service station or the ground cloud center, realizing the joint optimization of task processing time delay and energy consumption; The task processing location decision model in step S2 comprises: according to the task data volume, the bandwidth between the satellite and the ground edge service station, and the bandwidth between the satellite and the ground cloud center, the transmission time of the task data to the ground edge service station and the ground cloud center is calculated respectively; according to the computing capacity of the satellite end, the ground edge service station and the ground cloud center respectively, the computing processing time of the task in each node is estimated; the total completion time of the task in the satellite end, in the ground edge service station and in the ground cloud center is obtained by comprehensively considering the transmission time and the computing processing time; the total completion time of each processing path is compared with the maximum allowed response time of the task, and a feasible processing location set is preliminarily determined; a weighted cost function is introduced to calculate the comprehensive cost of each feasible processing location, and the node with the minimum value of the weighted cost function is selected as the final execution place of the task; S3. Constructing a geoscience space intelligent learning framework, performing offline training and optimization of the deep learning model in the ground cloud center, generating a lightweight model through model pruning and quantization, and delivering it to the satellite end for online inference and data collection, forming a closed-loop learning process of ground training-satellite reasoning-data back transmission-model iteration; S4. Introducing an offline-online collaborative processing mode, preferentially performing online processing on the satellite or the ground edge service station for high real-time tasks, and performing batch downloading for non-real-time tasks, dynamically adjusting the task allocation strategy according to the link state and node load, and improving the response efficiency and resource utilization of the system. 2.The edge-computing-based star-ground cooperative processing method according to claim 1, wherein, In step S2, the weighted cost function is a linear weighted sum of time delay cost and energy consumption cost. 3.The edge-computing-based star-ground cooperative processing method according to claim 1, wherein, The geoscience space intelligent learning framework in step S3 comprises: receiving and integrating the remote sensing data set periodically returned by the satellite in the ground cloud center, the data set comprising input samples and their corresponding labels; selecting the structure and size of the neural network model according to the computing capacity and storage capacity constraints of the satellite end, to ensure that the model meets the resource limitations of on-board deployment; using an iterative algorithm based on small batch random gradient descent to optimize the model parameters, and training the model by minimizing an empirical loss function containing a regularization term; in the training process, an independent validation data set is used to evaluate and monitor the model performance, and the hyperparameters are adjusted according to the evaluation results to obtain a trained initial model; performing model optimization on the trained initial model, including structured pruning based on parameter importance evaluation to remove redundant connections, and low-bit-width conversion based on weight quantization to reduce model storage and calculation overhead, thereby generating a lightweight model; downloading the lightweight model to the satellite end through a satellite-ground link; the satellite end uses the downloaded model to perform online inference on real-time collected data, and returns the inference results and newly collected data to the ground cloud center at a preset period; the ground cloud center updates the training data set using the returned data, and starts a new round of model training and optimization, forming a closed-loop iterative learning process. 4.The edge computing based star-ground cooperative processing method according to claim 3, wherein, The structured pruning includes: identifying redundant connections in the model through an importance evaluation function, and pruning according to a preset retention ratio or threshold to retain connections with high importance; calculating the compression rate of the pruned model to evaluate the impact of pruning on model performance. 5.The edge computing based star-ground cooperative processing method according to claim 3, wherein, The weight quantization includes: mapping floating-point weights to low-bit-width fixed-point integers, and calculating quantization error according to weight range and quantization bit width; balancing model accuracy and storage overhead by adjusting quantization bit width. 6.The edge computing based star-ground cooperative processing method according to claim 3, wherein, The online inference stage at the satellite end includes: estimating inference calculation complexity according to the number of model layers and parameter size; estimating inference time in combination with on-board hardware computing capacity; collecting new data at a preset period and returning it to the ground cloud center for model iterative update. 7.The edge-computing-based star-ground cooperative processing method according to claim 1, wherein, The offline-online collaborative processing mode in step S4 includes: defining a binary decision variable for each task to represent the decision state of whether the task is assigned to online processing or offline caching; building a scheduling optimization model with the goal of minimizing total system processing time; setting constraints for the optimization model, including: the completion time of each online processing task meets the maximum allowed response time requirement, and the total amount of data cache occupied by all online processing tasks does not exceed the upper limit of satellite end cache capacity; dynamically solving the optimization model according to the real-time monitored satellite-ground link state and node load state, updating the binary decision variable to dynamically adjust the task processing strategy.

8. The edge computing-based satellite-ground collaborative processing method of claim 7, wherein: The objective function of the scheduling optimization model takes into account the time cost of online processing at the satellite end and the time cost of caching tasks for download to the ground cloud center for processing, including ground cloud computing time and satellite-ground data transmission time; The constraints ensure that high real-time tasks are quickly responded to through online processing, while managing satellite end storage resource usage through cache capacity constraints; By dynamically adjusting the binary decision variable, intelligent switching and adaptive allocation between online processing and offline caching download modes are achieved.

9. An edge computing based space-ground collaborative processing system, configured to implement the edge computing based space-ground collaborative processing method according to any one of claims 1 to 8. The system includes: a satellite terminal layer including multiple satellite devices with edge computing capabilities for data collection, preliminary processing and online inference; a ground edge layer including multiple distributed ground edge service stations for receiving satellite data, providing edge computing services and communicating with the satellite end and the ground cloud center; a cloud computing layer including a ground cloud center for mass data storage, deep learning model training and system resource coordination; The task scheduling module is configured to dynamically allocate tasks to the satellite end, the ground edge service station or the ground cloud center according to the task characteristics and the system state based on real-time sensing of the resource state and the network environment of each node. The intelligent learning module is configured to implement training, optimization, distribution and iterative updating of the model, and support lightweight inference and continuous learning of the satellite end. The offline-online processing control module is configured to dynamically select an online processing or cache download strategy according to the task real-time requirement and the link state, and improve system resource utilization and response efficiency.

10. The edge computing-based satellite-ground cooperative processing system of claim 9, wherein: The task scheduling module is configured to dynamically select a task processing location based on a weighted cost function according to the real-time resource state and the link bandwidth of each node in the satellite end, the ground edge layer and the cloud computing layer, based on the task data volume, the calculation complexity and the maximum allowed response time, and to realize elastic distribution and migration of the task between the satellite and the ground. The intelligent learning module is configured to realize ground centralized training, lightweight optimization, satellite-ground safe distribution, satellite end online inference and periodic data backhaul of a deep learning model in cooperation with the ground cloud center and the satellite terminal layer, and to form a closed-loop iterative learning process. The offline-online processing control module is configured to dynamically decide a task execution strategy according to the task real-time requirement and the link load state in cooperation with the task scheduling module and the satellite-ground communication link, to control online processing of high real-time tasks in the satellite or the ground edge service station, and to cache and batch download non-real-time tasks to the ground cloud center for processing.

Images