Performance measurement method, device and equipment of satellite-ground cooperative fine adjustment system and medium

Abstract

The application provides a performance measurement method, device and equipment of a satellite-ground cooperative fine-tuning system and a medium, and relates to the technical field of satellite computing. The method comprises the following steps: constructing a satellite-ground coordination fine-tuning system, including a plurality of satellites and a ground station; configuring a local training module on the satellites and a satellite selection module and an aggregation module on the ground station; measuring the convergence of a satellite-borne model of the satellite-ground coordination fine-tuning system under a plurality of unstable link conditions based on TCP protocol and UDP protocol respectively, obtaining a first measurement result for representing a target protocol adapted to each unstable link condition; and measuring the convergence of the satellite-borne model of the satellite-ground coordination fine-tuning system under the plurality of unstable link conditions based on a plurality of priority transmission strategies respectively, obtaining a second measurement result for representing a target priority transmission strategy adapted to the unstable link condition under the target protocol adapted to each unstable link condition.

Description

Technical Field

[0001] This application relates to the field of satellite computing technology, and in particular to a performance measurement method, device, equipment and medium for a satellite-ground coordinated fine-tuning system. Background Technology

[0002] With the continuous advancement of satellite technology, an increasing number of satellites are being deployed to low Earth orbit (LEO). These satellites, with their unique architecture and orbital positions, focus on high-precision image acquisition. Simultaneously, with the improvement of satellite hardware computing power, more and more intelligent models are being deployed in orbit, greatly enhancing on-orbit service capabilities. Although intelligent models on-orbit satellites have the potential to enhance on-orbit services, due to limited satellite resources, many satellite-based machine learning tasks still rely on ground computing resources for collaboration. LEO satellites transmit data to ground stations via satellite-to-ground links, where model fine-tuning is performed within the powerful computing clusters of the ground stations. However, with increasing privacy protection requirements and rising satellite-to-ground link costs, the limitations of this traditional method are becoming increasingly apparent. Federated learning, due to its advantages in data privacy protection and distributed collaboration, has become a promising solution. Specifically, multiple satellites, acting as users in federated learning, transmit model parameters to the ground station for aggregation and updates after completing local model fine-tuning. This method alleviates privacy issues to some extent but fails to completely solve the communication overhead problem caused by satellite-to-ground links.

[0003] More importantly, most current research assumes link stability and focuses on optimizing federated aggregation algorithms or adjusting onboard model fine-tuning algorithms. These studies contribute to improving the efficiency of satellite-ground collaborative fine-tuning under stable links. However, the instability of satellite-ground links remains a significant challenge. Due to the high-speed motion of satellites and the complex space environment, satellite-ground links often exhibit high packet loss rates and long delays. Although some studies have explored this issue, most remain at the theoretical analysis stage, failing to conduct in-depth quantitative analysis of link conditions and their specific impact on system fine-tuning through experiments, particularly regarding performance disadvantages under high packet loss and long delay environments (such as convergence delay multiples and additional communication overhead ratios) and performance differences under different communication protocols.

[0004] Different communication protocols (such as TCP, UDP, and others) can significantly impact the performance of a space-ground collaborative fine-tuning system. Specifically, the choice of protocol directly affects data transmission latency, packet loss rate, and network bandwidth utilization, thus affecting the efficiency of model fine-tuning and the final training results. In environments with high packet loss rates and long latency, UDP, due to its low latency, can effectively reduce communication overhead and improve transmission efficiency. While TCP has advantages in data reliability, its retransmission mechanism and congestion control can significantly reduce efficiency in unstable link environments. Therefore, selecting an appropriate protocol is crucial to the overall performance of the space-ground collaborative fine-tuning system.

[0005] Therefore, related technologies typically lack systematic analysis of the performance of different communication protocols (such as TCP and UDP) under various unstable satellite-to-ground link environments, especially in scenarios with high packet loss rates and high latency. The choice of communication protocol directly impacts data transmission efficiency and model fine-tuning results, but existing research lacks quantitative data support, making it impossible to select suitable communication protocols under different unstable link conditions. Furthermore, many solutions are designed based on ideal link environments, failing to consider the volatility of satellite-to-ground links, such as dynamic changes in packet loss rates and communication interruptions. Insufficient adaptability to different link volatility scenarios makes it difficult to effectively address performance fluctuations in satellite-to-ground collaborative fine-tuning systems in practical deployments. In addition, balancing communication overhead and model accuracy remains a challenge in practical applications. Related technologies have failed to provide a method for dynamically selecting suitable parameter transmission strategies under different unstable link conditions to maximize communication efficiency. Summary of the Invention

[0006] In view of this, embodiments of this application provide a performance measurement method, apparatus, device, and medium for a satellite-ground coordinated fine-tuning system, in order to overcome the above problems or at least partially solve the above problems.

[0007] The first aspect of this application provides a performance measurement method for a satellite-ground collaborative fine-tuning system, the method comprising:

[0008] A satellite-ground coordinated fine-tuning system is constructed, comprising multiple satellites and a ground station. Each satellite is equipped with a local training module for performing local training of the onboard model, generating fine-tuning parameters, and sending them to the ground station. The ground station is equipped with a satellite selection module and an aggregation module. The satellite selection module is used to select satellites to participate in parameter aggregation, and the aggregation module is used to aggregate the fine-tuning parameters of the selected satellites to generate global model parameters.

[0009] Based on TCP and UDP protocols respectively, the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions is measured to obtain a first measurement result. The first measurement result is used to represent the target protocol adapted to each unstable link condition.

[0010] Based on the first measurement result, the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions is measured according to multiple priority transmission strategies, and a second measurement result is obtained. The priority transmission strategy is used to indicate the order in which the fine-tuning parameters are transmitted, and the second measurement result is used to represent the target priority transmission strategy adapted to each unstable link condition under the target protocol.

[0011] Optionally, the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions is measured based on TCP and UDP protocols respectively, to obtain the first measurement result, including:

[0012] Network simulation tools are used to simulate various unstable link conditions, including packet loss rate, latency, and network fluctuations.

[0013] Initialize the onboard model and training rounds;

[0014] Under different unstable link conditions, communication links between the satellite and the ground station are established using TCP and UDP protocols respectively;

[0015] The spaceborne model is trained, and the convergence status of the training is recorded. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold.

[0016] Based on the convergence of each training round, the target protocol for each unstable link condition is determined from the TCP and UDP protocols.

[0017] Optionally, based on the first measurement result, the convergence of the onboard model of the satellite-ground coordinated fine-tuning system under various unstable link conditions is measured according to multiple priority transmission strategies to obtain the second measurement result, including:

[0018] Network simulation tools are used to simulate various unstable link conditions, including packet loss rate, latency, and network fluctuations.

[0019] Initialize the onboard model and training rounds;

[0020] Under each unstable link condition, a communication link between the satellite and the ground station is established according to the target protocol adapted to that unstable link condition.

[0021] The spaceborne model is trained based on various priority transmission strategies, and the convergence status of the training is recorded. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold.

[0022] Based on the convergence of the training, the priority transmission strategy with the highest transmission efficiency among the various priority transmission strategies is selected as the target priority transmission strategy for adapting to this unstable link condition.

[0023] Optionally, the plurality of priority transmission strategies include: no-priority transmission strategy, gradient-assigned priority transmission strategy, and dynamic priority transmission strategy;

[0024] The no-priority transmission strategy means that the fine-tuning parameters of each satellite after local training are transmitted in sequence.

[0025] The gradient assignment priority transmission strategy means that the transmission priority of the fine-tuning parameters after local training of each satellite is determined according to the magnitude of the gradient of the fine-tuning parameters.

[0026] The dynamic priority transmission strategy means that the transmission priority is dynamically determined based on the contribution of each fine-tuning parameter to the convergence speed in each round of training after local training of each satellite.

[0027] Optionally, the method further includes:

[0028] When the target protocol is UDP, the onboard model is trained based on UDP, and the convergence status of the training is recorded. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold.

[0029] Based on whether the convergence of the training reaches the expected convergence, it can be determined whether the satellite-ground collaborative fine-tuning system should adopt a retransmission mechanism in actual application.

[0030] When the training convergence reaches the expected first target convergence, it is determined that the retransmission mechanism will not be used on the satellite-ground collaborative fine-tuning system.

[0031] When the training convergence reaches the expected second target convergence, the retransmission mechanism is determined to be used on the satellite-ground collaborative fine-tuning system.

[0032] Wherein, the convergence of the first objective is better than that of the second objective; the retransmission mechanism means that after the satellite transmits the fine-tuning parameters of the selected satellites participating in parameter aggregation to the ground station, and when the communication resources of the communication link between the satellite and the ground station are not exhausted, the fine-tuning parameters of the satellites participating in parameter aggregation that were not successfully transmitted will be retransmitted.

[0033] Optionally, the method further includes:

[0034] Under the current unstable link conditions, the satellite selection module of the ground station selects multiple satellites as target satellites from multiple satellites within the connection window according to preset rules; the preset rules are set based on the current status of each satellite, the quality and correlation of the local data collected by the satellites.

[0035] The multiple target satellites establish communication links with the ground station through a communication module, according to the target protocol adapted to the current unstable link conditions. According to the target priority transmission strategy adapted to the current unstable link conditions, the fine-tuning parameters corresponding to each of the multiple target satellites are transmitted to the ground station. The fine-tuning parameters are obtained by the target satellites through a local training module based on local data to train the onboard model deployed locally on the target satellites.

[0036] The ground station uses an aggregation module to aggregate multiple received fine-tuning parameters to obtain the global model parameters for this round.

[0037] The ground station transmits the global model parameters to the multiple satellites through a communication link established between the ground station and the target protocol adapted to the current unstable link conditions. This enables the multiple satellites to update the current model parameters of their locally deployed onboard models based on the global model parameters. The satellite selection module then selects new target satellites from the multiple satellites to participate in the next round of local training until the onboard model training is completed.

[0038] Specifically, when the spaceborne model is used to implement disaster navigation, the local data is disaster information data collected by the satellite; when the spaceborne model is used to implement Earth observation, the local data is Earth observation data collected by the satellite; and when the spaceborne model is used to implement climate monitoring, the local data is climate monitoring data collected by the satellite.

[0039] Optionally, the ground station, through an aggregation module, aggregates multiple received fine-tuning parameters to obtain the global model parameters for this round, including:

[0040] When the fine-tuning parameters of any satellite among the selected satellites participating in parameter aggregation fail to be transmitted, the fine-tuning parameters of multiple satellites that successfully transmitted and the fine-tuning parameters of the target satellite that failed to transmit in the previous round are aggregated to obtain the global model parameters for this round.

[0041] When the fine-tuning parameters of all selected satellites participating in parameter aggregation are successfully transmitted, the fine-tuning parameters of all selected satellites participating in parameter aggregation are aggregated to obtain the global model parameters for this round.

[0042] A second aspect of this application provides a performance measurement device for a satellite-ground collaborative fine-tuning system, the device comprising:

[0043] A construction module is used to build a satellite-ground coordinated fine-tuning system, which includes multiple satellites and a ground station. The satellites are configured with a local training module to perform local training of the onboard model, generate fine-tuning parameters, and send them to the ground station. The ground station is configured with a satellite selection module and an aggregation module. The satellite selection module is used to select satellites to participate in parameter aggregation, and the aggregation module is used to aggregate the fine-tuning parameters of the selected satellites to generate global model parameters.

[0044] The first measurement module is used to measure the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions based on TCP and UDP protocols respectively, and obtain the first measurement result. The first measurement result is used to represent the target protocol adapted to each unstable link condition.

[0045] The second measurement module is used to measure the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions based on the first measurement result and various priority transmission strategies, respectively, to obtain the second measurement result; the priority transmission strategy is used to indicate the order of transmitting the fine-tuning parameters, and the second measurement result is used to represent the target priority transmission strategy adapted to each unstable link condition under the target protocol.

[0046] A third aspect of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method as described in the first aspect.

[0047] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0048] The beneficial effects of this application are:

[0049] This application provides a performance measurement method, apparatus, device, and medium for a satellite-ground collaborative fine-tuning system. Through the technical solution of this application, the collaborative working mode between the satellite and the ground station is clarified by constructing a satellite-ground collaborative fine-tuning system, providing a basic architecture for subsequent performance measurements. Secondly, by performing measurements based on TCP and UDP protocols respectively, the performance differences of different protocols under unstable link conditions are quantitatively analyzed, providing data support for selecting a suitable transmission protocol. Finally, by introducing multiple priority transmission strategies for measurement, the data transmission process of fine-tuning parameters is further optimized, reducing communication overhead, improving communication efficiency, and increasing the convergence speed and accuracy of the model. This enables the system to achieve efficient and stable model fine-tuning under complex and variable satellite-ground link conditions, significantly improving the overall performance of the system. Attached Figure Description

[0050] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0051] To more clearly illustrate the technical solution of this application, the drawings used in the description of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of a performance measurement method for a satellite-ground collaborative fine-tuning system provided in one embodiment of this application;

[0053] Figure 2 This is a system configuration diagram of a satellite-ground coordination fine-tuning system provided in an embodiment of this application;

[0054] Figure 3 This is a schematic diagram of the framework of a performance measurement device for a satellite-ground collaborative fine-tuning system provided in one embodiment of this application. Detailed Implementation

[0055] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0056] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] TCP (Transmission Control Protocol) is a connection-oriented, reliable, byte-stream-based transport layer communication protocol.

[0058] UDP: User Datagram Protocol. UDP uses IP as its underlying protocol and provides applications with a minimal protocol mechanism to send messages to other programs. Its main characteristics are connectionless, no guaranteed reliable transmission, and message-oriented nature.

[0059] FL: Federated Learning is a distributed machine learning framework designed to help multiple organizations share knowledge and perform machine learning modeling while adhering to user privacy protection, data security, and government regulations.

[0060] Figure 1 This is a flowchart illustrating a performance measurement method for a satellite-ground coordinated fine-tuning system according to an embodiment of this application. (Refer to...) Figure 1 One embodiment of this application provides a performance measurement method for a satellite-ground coordinated fine-tuning system, the method comprising:

[0061] Step S11: Construct a satellite-ground coordinated fine-tuning system, which includes multiple satellites and a ground station; each satellite is equipped with a local training module for performing local training of the onboard model, generating fine-tuning parameters and sending them to the ground station; the ground station is equipped with a satellite selection module and an aggregation module, the satellite selection module for selecting satellites to participate in parameter aggregation, and the aggregation module for aggregating the fine-tuning parameters of the selected satellites to generate global model parameters.

[0062] In this embodiment, the satellite-ground collaborative fine-tuning system consists of two parts: a satellite constellation (containing multiple satellites) and a ground station. This system achieves efficient data transmission and model fine-tuning between the satellites and the ground station through the collaborative work of multiple modules. The core objective of the system is to optimize module design under unstable satellite-ground link conditions, ensuring stability and efficiency during the model fine-tuning process.

[0063] In constructing the space-ground coordinated fine-tuning system, we first consider a set of low Earth orbit satellites and ground stations g for Earth observation. A continuous clock time t and a discrete time index are introduced. i

[0064] ∈0,1,2,..., each adjacent time index has a time interval of τ. From... i τ to ( i +1) The time interval of τ is expressed as [ i When satellite s∈S at any time t∈[ i If satellite s is captured by ground station g∈G, a communication link will be established between satellite s and ground station g. Satellite s is defined as the time interval [ i In the context of satellites connected to ground station G, satellite s can utilize the communication resources during the connection establishment process via the communication link. When satellite s moves out of the signal coverage area of ​​ground station g, the communication link is interrupted, and satellite s is no longer connected to ground station g. Satellite s can no longer utilize the communication resources during the connection establishment process via the communication link. The duration of this communication link connection is defined as the connection window, denoted as τd, and τd ≤ τ. Furthermore, it is assumed that all ground stations operate as a powerful computing cluster, without considering transmission delays between these geographically distributed ground stations. This assumption is theoretically reasonable because the structural connections between individual ground stations are always well-configured and much faster than satellite-to-ground links. Therefore, a satellite connected to any ground station can be considered connected to a cluster with all ground stations G.

[0065] Figure 2 This is a system configuration diagram of a satellite-to-ground coordination fine-tuning system provided in an embodiment of this application, with reference to... Figure 2 The satellite-ground coordinated fine-tuning system of this application consists of two parts: a satellite constellation and a ground station. Each satellite is equipped with an onboard model to be trained. The system achieves efficient data transmission and model fine-tuning through the collaborative work of multiple modules. The system module structure of this satellite-ground coordinated fine-tuning system includes four modules: a local training module, a communication module (divided into satellite end and ground station end), a satellite selection module, and an aggregation module.

[0066] The satellite constellation consists of two main modules: a local training module and a communication module (satellite end).

[0067] Local Training Module: Each satellite in the constellation uses its own locally collected data, such as terrain and vegetation data from Earth observation, to train its onboard model locally. Local training of the onboard model on the satellite significantly reduces the need for transmitting raw data to the ground station, saving bandwidth and time, and improving fine-tuning efficiency. The fine-tuning parameters (such as gradients and weights) of the onboard model obtained through local training are then transmitted to the ground station for aggregation after local training.

[0068] Communication Module (Satellite End): The satellite transmits updated fine-tuning parameters (such as fine-tuned gradients and weights) to the ground station for aggregation via the communication module. The communication module uses an adapted communication protocol (such as TCP or UDP) to transmit the fine-tuning parameters, ensuring that data transmission between the satellite and the ground station is both efficient and stable.

[0069] Ground station section: The ground station mainly consists of a satellite selection module, an aggregation module, and a communication module (ground station end).

[0070] Satellite Selection Module: The ground station uses the satellite selection module to select satellites to participate in each round of fine-tuning parameter aggregation according to preset rules. These preset rules can be based on the satellite's current status (such as energy level, computing power, etc.) or the quality and relevance of the data collected by the satellite. For example, if certain satellites collect richer or higher-quality local data in a specific area, they may be prioritized for fine-tuning parameter aggregation of the onboard model.

[0071] Aggregation Module: After the ground station receives fine-tuning parameters from multiple satellites, the aggregation module is responsible for aggregating these parameters and generating global model parameters. During the aggregation process, the aggregation module can consider factors such as data quality and computing power of different satellites, and perform a weighted average of the fine-tuning parameters to generate a more representative global model parameter.

[0072] Communication Module (Ground Station): Since the ground station and satellites need to perform multiple rounds of collaborative fine-tuning, the system also includes a feedback mechanism. After aggregating multiple fine-tuning parameters to obtain global model parameters in each round, these global model parameters need to be sent to each satellite via the communication module (ground station). This allows each satellite to update the parameters of its locally deployed onboard model based on the global model parameters and perform the next round of local training. In this way, the satellite and ground station achieve multi-round federated learning collaboration, ultimately achieving the goal of optimizing the global model (i.e., the onboard model obtained through multiple rounds of collaborative fine-tuning).

[0073] Step S12: Based on TCP and UDP protocols respectively, measure the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions to obtain a first measurement result. The first measurement result is used to represent the target protocol adapted to each unstable link condition.

[0074] In this embodiment, we consider that using different communication protocols to establish communication links between the satellite and the ground station will significantly affect the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions (such as different packet loss rates, long delays, and network fluctuations). Therefore, in order to determine which protocol is more suitable for system operation under each unstable link condition, it is necessary to measure the system performance under various unstable link conditions based on TCP and UDP protocols respectively.

[0075] TCP is a connection-oriented, reliable transport layer protocol that guarantees data integrity and order. However, its retransmission mechanism and congestion control may reduce transmission efficiency in environments with high packet loss rates and long delays.

[0076] UDP is a connectionless protocol that does not guarantee reliability. Its advantages are low latency and high throughput, but in environments with high packet loss rates, additional mechanisms may be needed to ensure data integrity.

[0077] By conducting experiments using these two protocols under different link conditions (such as different combinations of packet loss rates and delays), corresponding measurement results, known as the first measurement results, can be obtained. These first measurement results represent the target protocol (UDP or TCP) adapted for each unstable link condition. Quantitative analysis of the system performance of different protocols under unstable link conditions provides fundamental data support for subsequent priority transmission strategy selection. This allows for direct selection of the appropriate target protocol to establish the communication link between the satellite and ground station under the current unstable link condition, based on the first measurement results, during subsequent actual satellite-ground collaborative fine-tuning. This improves system performance and enables the onboard model to achieve optimal convergence, enhancing both convergence speed and model accuracy.

[0078] Step S13: Based on the first measurement result, the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions is measured according to multiple priority transmission strategies to obtain a second measurement result; the priority transmission strategy is used to indicate the order of transmitting the fine-tuning parameters, and the second measurement result is used to represent the target priority transmission strategy adapted to each unstable link condition under the target protocol.

[0079] In this embodiment, after determining the communication protocols suitable for different unstable link conditions, in order to further optimize the system performance, this embodiment takes into account the limited communication resources of the satellite-to-ground link. Therefore, it is necessary to further determine how to efficiently transmit the fine-tuning parameters of the satellites participating in the aggregation. Based on this, we believe that when the communication link between the satellite and the ground is unstable, the system performance can be significantly improved by reasonably arranging the transmission order of the fine-tuning parameters. Higher system performance can lead to better convergence speed and model accuracy of the onboard model.

[0080] Therefore, in this step, multiple pre-set priority transmission strategies are used to measure the convergence of the satellite-ground coordinated fine-tuning system under various unstable link conditions, and a second measurement result is obtained.

[0081] The priority transmission strategy refers to determining the order in which fine-tuning parameters of each satellite are transmitted via the satellite-to-ground communication link, based on factors such as the importance of the parameters or their contribution to model convergence. For example, a priority transmission strategy can be defined based on criteria such as parameter gradient magnitude and parameter sensitivity. This approach optimizes data transmission, reduces unnecessary communication overhead, and improves the convergence speed and accuracy of the onboard model. The second measurement result represents the priority transmission strategy adapted to each unstable link condition among various priority transmission strategies when establishing a satellite-to-ground communication link using a target protocol adapted to each unstable link condition; this is the target priority transmission strategy. Verifying the effectiveness of different priority transmission strategies under different link conditions provides data support for selecting the optimal transmission strategy in practical applications.

[0082] By combining the two measurement processes mentioned above, we can obtain the target protocol and target priority transmission strategy adapted to each unstable link condition. In the subsequent practical application of the system, we can directly select the target protocol and target priority transmission strategy adapted to the current unstable link condition based on the measurement results to carry out the process of satellite-ground collaborative fine-tuning of the satellite model, so as to maximize the performance of the system and make the convergence speed and model accuracy of the satellite model better.

[0083] The technical solutions described above achieve performance optimization of the satellite-ground collaborative fine-tuning system under unstable link conditions. First, by constructing the satellite-ground collaborative fine-tuning system, the collaborative working mode between the satellite and the ground station is clarified, providing the basic architecture for subsequent performance measurements. Second, by conducting measurements based on TCP and UDP protocols respectively, the performance differences of different protocols under unstable link conditions are quantitatively analyzed, providing data support for selecting a suitable transmission protocol. Finally, by introducing multiple priority transmission strategies for measurement, the data transmission process of fine-tuning parameters is further optimized, communication overhead is reduced, and the convergence speed and accuracy of the model are improved. This enables the system to achieve efficient and stable model fine-tuning under complex and variable satellite-ground link conditions, significantly improving the overall performance of the system.

[0084] In conjunction with the technical solutions of the above embodiments, an embodiment of this application also provides another performance measurement method for a satellite-ground coordinated fine-tuning system. In this method, step S12, "measuring the convergence of the onboard model of the satellite-ground coordinated fine-tuning system under various unstable link conditions based on TCP and UDP protocols respectively, and obtaining a first measurement result," specifically includes steps S12-1 to S12-5:

[0085] Step S12-1: Use a network simulation tool to simulate various unstable link conditions, including packet loss rate, latency, and network fluctuations.

[0086] In this embodiment, in the satellite-ground collaborative fine-tuning system, the communication link between the satellite and the ground station is often unstable due to various factors, such as the satellite's high-speed motion, atmospheric interference, and the complexity of the space environment. This instability manifests primarily as high packet loss rates, long delays, and intermittent link interruptions. To accurately evaluate the performance of different communication protocols under these complex and unstable link conditions, network simulation tools are needed to simulate various unstable link conditions. This involves simulating different instabilities that may occur in the satellite-ground communication link, including packet loss rates (e.g., 0%, 5%, 10%, 20%) and delays (e.g., 100ms, 300ms, 500ms), and setting different network volatility levels. Training experiments are conducted for each condition. Using network simulation tools (such as NS3), the parameters of these conditions can be precisely set and adjusted, providing a controllable and diverse testing environment for subsequent experiments. This step is fundamental to the entire performance measurement method, ensuring the reliability and repeatability of the measurement results.

[0087] Step S12-2: Initialize the onboard model and training rounds.

[0088] In this embodiment, the onboard model needs to be initialized before performance measurement. The onboard model refers to a machine learning model deployed on a satellite, used for local training and fine-tuning of data collected by the satellite. The initialization process includes setting the model's initial parameters, selecting a suitable training algorithm, and determining the number of training epochs. The number of training epochs refers to the number of iterations the model performs during training, typically determined by the model's complexity and the size of the dataset. By initializing the onboard model and the number of training epochs, it is ensured that experiments are conducted under the same starting conditions, making performance comparisons under different protocols and unstable link conditions comparable.

[0089] For example, the choice of onboard model can use classic deep learning models (such as convolutional neural networks CNN), which are suitable for tasks such as image recognition or data analysis. The number of training rounds can be set, such as 100 rounds, and the loss value and accuracy of the onboard model in each round can be recorded.

[0090] Step S12-3: Under different unstable link conditions, establish communication links between the satellite and the ground station using TCP and UDP protocols respectively.

[0091] In this embodiment, under simulated unstable link conditions, communication links between the satellite and the ground station are established using both TCP and UDP protocols. By establishing communication links using these two protocols under the same unstable link conditions, the system performance of the two protocols under different unstable link conditions can be compared and analyzed, thereby providing data support for selecting an appropriate target protocol in subsequent practical applications.

[0092] Step S12-4: Train the spaceborne model and record the convergence status of the training. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold.

[0093] In this embodiment, after establishing a communication link, different protocols will be used to collaboratively fine-tune the onboard model under various unstable link conditions. This includes local training on the satellite and parameter aggregation at the ground station, and the convergence status during the fine-tuning process will be recorded. Convergence status includes training time, convergence time, loss value (e.g., mean squared error, MSE), and accuracy. Training time represents the total time required for training, estimated by using the time for each training round, with the time unit being seconds. Convergence time refers to the time required for training to reach a preset error threshold (e.g., MSE < 0.01), which is determined based on the loss value of each training round, the current training round, the total number of training rounds, and the error threshold. The loss value measures the model's error during training, while accuracy reflects the model's ability to classify or predict data. By recording these key indicators, the performance of different protocols under different link conditions can be comprehensively evaluated, particularly their impact on model convergence speed and accuracy.

[0094] For example, the formula for calculating the convergence time is:

[0095] ;in, This represents the loss in each training round, where t is the current training round and N is the total number of rounds. This is the error threshold.

[0096] For example, the formula for calculating training time is:

[0097] ;in, This represents the training time for round t.

[0098] Step S12-5: Based on the convergence of each training round, determine the target protocol for each unstable link condition from the TCP and UDP protocols.

[0099] In this embodiment, after training and performance recording of different protocols under different link conditions are completed, the protocol adapted to each unstable link condition is determined from the TCP and UDP protocols as the target protocol based on the convergence of each training round.

[0100] For example, if, under a certain link condition, the convergence time of the UDP protocol is significantly shorter than that of the TCP protocol, and it also exhibits better performance in terms of loss value and accuracy, then the UDP protocol can be considered the target protocol under that link condition. This step provides data support for the selection of communication protocols under different link conditions in the satellite-ground collaborative fine-tuning system, thereby optimizing the overall performance of the system.

[0101] The technical solutions described above utilize network simulation tools to model various unstable link conditions, providing a controllable and diverse testing environment for the experiments and ensuring the reliability and repeatability of the experimental results. Secondly, by initializing the onboard model and training rounds, a unified benchmark is provided for subsequent training and performance evaluation. Next, communication links are established using TCP and UDP protocols under different unstable link conditions, and the onboard model is trained, with the convergence status of the training recorded. This process, by comparing and analyzing the performance of different protocols under different link conditions, provides data support for selecting a suitable protocol. Finally, based on the convergence status of each training round, the target protocol suitable for each unstable link condition is determined.

[0102] In conjunction with the technical solutions of the above embodiments, an embodiment of this application also provides another performance measurement method for a satellite-ground coordinated fine-tuning system. In this method, step S13, "based on the first measurement result, measuring the convergence of the onboard model of the satellite-ground coordinated fine-tuning system under various unstable link conditions based on multiple priority transmission strategies, and obtaining the second measurement result," specifically includes steps S13-1 to S13-5:

[0103] Step S13-1 involves using a network simulation tool to simulate various unstable link conditions, including packet loss rate, latency, and network fluctuations. In this embodiment, this step is the same as step S12-1 described above, and will not be repeated here.

[0104] Step S13-2: Initialize the onboard model and training rounds. In this embodiment, this step refers to the content of step S12-2 above, and will not be repeated here.

[0105] Step S13-3: Under each unstable link condition, establish a communication link between the satellite and the ground station according to the target protocol adapted to that unstable link condition. In this embodiment, this step refers to the content of step S12-3 above, and will not be repeated here.

[0106] Step S13-4: Train the spaceborne model based on multiple priority transmission strategies, and record the convergence status of the training. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold.

[0107] In this embodiment, after establishing a communication link based on the target protocol, the onboard model is trained using various priority transmission strategies, and the convergence status of the training is recorded. The priority transmission strategy determines the order in which fine-tuning parameters are transmitted based on their importance or contribution to model convergence. Furthermore, the difference between the error values ​​from the training and validation processes is used to measure the impact of different priority transmission strategies on convergence speed and accuracy.

[0108] The difference between the error values ​​from the training process (obtained using the training dataset from the local dataset) and the error values ​​from the validation process (obtained using the validation dataset from the local dataset) is used to measure the impact of different priority transfer strategies on convergence speed and accuracy. The specific process is as follows:

[0109] ;

[0110] Here, Convergence difference indicates a difference. This represents the error value during the training process. This represents the error value during the verification process.

[0111] When the difference between training error and validation error is small and both tend to stabilize, it indicates that the current onboard model has good generalization ability and convergence performance. When the difference between training error and validation error is large, it may indicate that the model performs well on the training data but poorly on new, unseen data (i.e., the validation dataset), which means that the onboard model is overfitting. The convergence status can be referred to in steps S12-4 above, and will not be repeated here.

[0112] Step S13-5: Based on the convergence of the training, the priority transmission strategy with the highest transmission efficiency among the various priority transmission strategies is selected as the target priority transmission strategy for adapting to this unstable link condition.

[0113] In this embodiment, after recording the model fine-tuning and measurement results of various priority transmission strategies under different link conditions, under the same unstable link condition, based on the convergence of each training round, the priority transmission strategy with the highest transmission efficiency is selected as the target priority transmission strategy for adapting to that unstable link condition. For example, if, under a certain unstable link condition, the gradient assignment priority transmission strategy results in the shortest convergence time and the highest accuracy, then this strategy can be considered the target priority transmission strategy under that link condition. This step provides data support for the selection of priority transmission strategies for the satellite-ground collaborative fine-tuning system under different link conditions, thereby further optimizing the overall performance of the system.

[0114] For example, the formula for calculating transmission efficiency is:

[0115] ;

[0116] Where M is the number of fine-tuning parameters transmitted. Let i be the size of the i-th fine-tuning parameter. This represents the total transmission time.

[0117] In conjunction with the technical solutions of the above embodiments, an embodiment of this application also provides another performance measurement method for a satellite-ground collaborative fine-tuning system. In this method, the multiple priority transmission strategies include: a no-priority transmission strategy, a gradient-assigned priority transmission strategy, and a dynamic priority transmission strategy.

[0118] The no-priority transmission strategy means that the fine-tuning parameters of each satellite after local training are transmitted in sequence.

[0119] Specifically, in the no-priority transmission strategy, the fine-tuned parameters trained locally on each satellite are transmitted sequentially. Under this strategy, all fine-tuned parameters are considered equally important, regardless of their contribution to model convergence.

[0120] The gradient assignment priority transmission strategy means that the transmission priority of the fine-tuning parameters after local training of each satellite is determined according to the magnitude of the gradient of the fine-tuning parameters.

[0121] Specifically, in the gradient assignment priority transmission strategy, the fine-tuning parameters trained locally on each satellite are prioritized for transmission based on the magnitude of their gradients. Parameters with larger gradient magnitudes typically have a greater impact on the model's convergence speed. Therefore, prioritizing the transmission of these parameters can accelerate model convergence, effectively utilizing limited communication resources and prioritizing the transmission of parameters more important to model convergence, thereby improving training efficiency.

[0122] The dynamic priority transmission strategy means that the transmission priority is dynamically determined based on the contribution of each fine-tuning parameter to the convergence speed in each round of training after local training of each satellite.

[0123] Specifically, in the dynamic priority transmission strategy, the fine-tuned parameters of each satellite after local training are dynamically prioritized based on the contribution of each fine-tuned parameter to the convergence speed in each training round. Unlike the gradient assignment priority transmission strategy, the dynamic priority transmission strategy can dynamically adjust the transmission order of parameters according to the current training state and link conditions.

[0124] For example, in some training epochs, certain fine-tuning parameters may contribute more to the convergence speed, so these fine-tuning parameters are transmitted first. In other training epochs, other fine-tuning parameters are more important, so other fine-tuning parameters are transmitted first. By dynamically adjusting the transmission priority, high communication efficiency and model convergence speed can be maintained under different link conditions.

[0125] The technical solutions described in the above embodiments introduce a no-priority transmission strategy, a gradient-assigned priority transmission strategy, and a dynamic priority transmission strategy, which can flexibly select the most suitable priority transmission strategy according to different link conditions and training requirements.

[0126] In conjunction with the technical solutions of the above embodiments, an embodiment of this application also provides another performance measurement method for a satellite-ground coordinated fine-tuning system. In this method, the method further includes steps S21 to S24:

[0127] Step S21: When the target protocol is UDP, the onboard model is trained based on the UDP protocol, and the convergence status of the training is recorded. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold.

[0128] In this embodiment, in the space-ground collaborative fine-tuning system, when the target protocol is UDP, the onboard model needs to be trained using UDP, and the convergence status during the training process needs to be recorded in detail. The convergence status is described above and will not be repeated here.

[0129] Step S22: Based on whether the convergence of the training has reached the expected convergence, determine whether the satellite-ground collaborative fine-tuning system should adopt a retransmission mechanism in actual application.

[0130] In this embodiment, after completing the UDP-based training and recording the convergence status, it is necessary to determine whether the satellite-ground collaborative fine-tuning system should employ a retransmission mechanism in practical applications, based on whether the training convergence has reached the expected target convergence. The introduction of a retransmission mechanism can improve the reliability of data transmission of fine-tuning parameters, ensuring that the fine-tuning parameters of the satellite model can reach the ground station completely and accurately, thereby improving the convergence speed and accuracy of the satellite model.

[0131] Step S23: When the training convergence reaches the expected first target convergence, determine that the retransmission mechanism will not be used on the satellite-ground collaborative fine-tuning system.

[0132] In this embodiment, when the convergence is good, that is, when the first target convergence is achieved, it means that the current performance of the spaceborne model is good. At this time, the impact of the fine-tuning parameters that failed to be transmitted on the spaceborne model is relatively small. Therefore, it is possible to choose not to retransmit these fine-tuning parameters that failed to be transmitted, but to prioritize the reduction of communication resources in the subsequent spaceborne model space-ground collaborative fine-tuning process.

[0133] Step S24: When the training convergence reaches the desired second target convergence, determine to adopt the retransmission mechanism on the satellite-ground collaborative fine-tuning system.

[0134] In this embodiment, when the convergence is relatively insufficient, that is, when the second target convergence is reached, it indicates that the current performance of the spaceborne model is poor. At this time, the fine-tuning parameters that failed to be transmitted have a relatively large impact on the spaceborne model. Therefore, it is necessary to retransmit these fine-tuning parameters that failed to be transmitted to ensure the quality of the subsequent spaceborne model's space-ground collaborative fine-tuning process.

[0135] Wherein, the convergence of the first objective is better than that of the second objective; the retransmission mechanism means that after the satellite transmits the fine-tuning parameters of the selected satellites participating in parameter aggregation to the ground station, and when the communication resources of the communication link between the satellite and the ground station are not exhausted, the fine-tuning parameters of the satellites participating in parameter aggregation that were not successfully transmitted will be retransmitted.

[0136] Suppose that in a data transmission, the satellite transmits 100 fine-tuning parameters to the ground station, but 10 parameters fail to transmit successfully. With unused communication resources, the satellite retransmits these 10 unsuccessful parameters to the ground station. This ensures that all parameters reach the ground station intact, thereby improving the model's convergence speed and accuracy.

[0137] For example, assuming an error threshold of 0.01, during training using the UDP protocol, we recorded the following data:

[0138] In the 10th round of training, the loss value was 0.02, and the accuracy rate was 95%.

[0139] During the 15th round of training, the loss value was 0.009, and the accuracy rate was 96%.

[0140] During the 20th round of training, the loss value was 0.005, and the accuracy rate was 97%.

[0141] During the 30th round of training, the loss value was 0.003 and the accuracy rate was 98%.

[0142] Based on this data, we can calculate that the loss value first falls below 0.01 during the 15th training round, and the convergence time is the time required to complete the 15th training round. In addition, we define two convergence scenarios. During each training round, we can determine whether to use a retransmission mechanism based on the recorded convergence data, where:

[0143] First objective convergence: loss value below 0.004, accuracy above 97%;

[0144] Second objective convergence: loss value below 0.01, accuracy above 95%.

[0145] In the 10th training round, the loss value was 0.02 and the accuracy was 95%, which met the second objective convergence condition. Under these conditions, higher performance requirements are needed, and a retransmission mechanism is required to further improve the reliability of the system and the convergence quality of the model.

[0146] During the 20th training round, the loss value was 0.005 and the accuracy was 97%, which met the first objective convergence condition. At this point, there was no need to use a retransmission mechanism. In other words, under the current link conditions and UDP protocol, the performance of the onboard model already met the basic requirements, and no additional retransmission mechanism was needed to improve reliability.

[0147] The technical solutions described above enable a dynamic decision on whether to employ a retransmission mechanism based on the training convergence status when using the UDP protocol, thereby achieving a balance between communication efficiency and reliability. This dynamic introduction of a retransmission mechanism improves data transmission reliability without significantly increasing communication overhead, ensuring that the onboard model parameters arrive at the ground station completely and accurately, thus improving the model's convergence speed and accuracy. Furthermore, dynamically selecting whether to employ a retransmission mechanism based on different convergence statuses allows the system to flexibly adjust transmission strategies under varying link conditions and performance requirements, optimizing the overall system performance.

[0148] In conjunction with the technical solutions of the above embodiments, one embodiment of this application also provides another performance measurement method for a satellite-ground coordinated fine-tuning system. In this method, the method further includes steps S31 to S34:

[0149] Step S31: Under the current unstable link conditions, the satellite selection module of the ground station selects multiple satellites as target satellites from multiple satellites in the connection window according to preset rules; the preset rules are set based on the current status of each satellite, the quality and correlation of the local data collected by the satellites.

[0150] In this embodiment, after the aforementioned measurement process is completed, the actual fine-tuning of the onboard model can be performed through the space-ground collaborative fine-tuning system. To ensure that the satellites participating in the model fine-tuning have high-quality data and sufficient resources, the satellite selection module at the ground station needs to select multiple satellites as target satellites from among the multiple satellites within the connection window. The selection process is based on preset rules, which comprehensively consider the current status of each satellite (such as energy level and computing power) as well as the quality and relevance of the local data collected by the satellite. Specifically, the current status of a satellite reflects its ability to participate in fine-tuning; for example, satellites with higher energy levels can perform computing tasks more stably. The quality and relevance of local data directly affect the effect of onboard model fine-tuning; high-quality data that is highly relevant to the task can improve the performance of the onboard model. In this way, the satellite selection module can screen out the most suitable satellites to participate in the current fine-tuning task, laying the foundation for subsequent efficient data transmission and model optimization.

[0151] In step S32, the multiple target satellites establish a communication link with the ground station through the communication module according to the target protocol adapted to the current unstable link conditions, and transmit the fine-tuning parameters corresponding to each of the multiple target satellites to the ground station according to the target priority transmission strategy adapted to the current unstable link conditions. The fine-tuning parameters are obtained by the target satellites through the local training module based on local data to train the onboard model deployed locally by the target satellites.

[0152] In this embodiment, after selecting target satellites, these satellites need to establish a communication link with the ground station through their communication modules, based on the first measurement results described above, to determine the target protocol suitable for the current unstable link conditions. They also need to determine the target priority transmission strategy suitable for the current unstable link conditions based on the second measurement results described above, and transmit their respective fine-tuning parameters to the ground station. The fine-tuning parameters are obtained by the target satellites through local training modules using local data to train the locally deployed onboard model. In this way, even under unstable link conditions, communication efficiency can be maximized, the impact of data transmission on model convergence can be reduced, and the efficiency and stability of the model fine-tuning process can be ensured.

[0153] In step S33, the ground station aggregates the received multiple fine-tuning parameters through the aggregation module to obtain the global model parameters for this round.

[0154] In this embodiment, after receiving fine-tuning parameters from multiple target satellites, the ground station aggregates these parameters through an aggregation module to generate the global model parameters for this round. The aggregation process integrates local model updates from different satellites into a single global model, thereby achieving collaborative optimization of model parameters. The aggregation module can perform a weighted average of the fine-tuning parameters from different satellites based on factors such as data quality and computing power, thus improving the generalization ability of the onboard model and ensuring its robustness across different satellite data. In this way, the ground station can effectively integrate dispersed satellite resources, enhance the collaborative working capability of the entire system, and provide optimized global model parameters for subsequent model updates and the next round of fine-tuning.

[0155] In step S34, the ground station transmits the global model parameters to the multiple satellites through a communication link established between the ground station and the target protocol adapted to the current unstable link conditions. This enables the multiple satellites to update the current model parameters of the locally deployed spaceborne model based on the global model parameters, and to reselect new target satellites from the multiple satellites through the satellite selection module to participate in the next round of local training until the spaceborne model training is completed.

[0156] Specifically, when the spaceborne model is used to implement disaster navigation, the local data is disaster information data collected by the satellite; when the spaceborne model is used to implement Earth observation, the local data is Earth observation data collected by the satellite; and when the spaceborne model is used to implement climate monitoring, the local data is climate monitoring data collected by the satellite.

[0157] In this embodiment, the ground station transmits the aggregated global model parameters to multiple satellites via a communication link. The satellites update the current model parameters of their locally deployed onboard models based on the global model parameters and then reselect new target satellites through a satellite selection module to participate in the next round of local training, until the onboard model training is complete. By transmitting the global model parameters back to the satellites, they can perform local updates based on the latest global model, thus better adapting to the global optimization objective in the next round of fine-tuning. Simultaneously, the satellite selection module reselects target satellites based on the updated model performance and satellite status, ensuring that each round of fine-tuning is performed with the optimal satellite combination. This iterative process not only improves the convergence speed of the onboard model but also enhances the system's adaptability, enabling efficient and stable model training under complex and unstable link conditions.

[0158] Furthermore, depending on the specific application function of the spaceborne model (such as disaster navigation, Earth observation, or climate monitoring), the type of local data will also vary, thus affecting the training direction and performance optimization goals of the model, resulting in a personalized spaceborne model.

[0159] For example, disaster navigation functionality uses local data collected by satellites, such as images of fire areas and monitoring data on flood extent; Earth observation functionality uses local data collected by satellites, such as topographic maps and vegetation cover; and climate monitoring functionality uses local data collected by satellites, such as atmospheric temperature and humidity. Different types of local data directly affect the satellite's local training process and the generation of fine-tuning parameters, thus impacting the performance of the final onboard model.

[0160] The technical solutions described above enable the efficient operation of the satellite-ground collaborative fine-tuning system under unstable link conditions. First, by optimizing the satellite selection process, it ensures that the satellites participating in the fine-tuning possess high-quality data and sufficient resources, thereby improving training efficiency. Second, through flexible communication strategies, the target protocol and target priority transmission strategy are dynamically selected based on link conditions and the first and second measurement results, ensuring efficient and reliable data transmission. Next, a global model parameter is generated through an aggregation module and transmitted back to the satellites, enabling multi-round iterative optimization to further improve model performance. Finally, through iterative iteration and dynamic satellite selection, the system can adaptively optimize the model training process, ensuring efficient and stable operation in complex and ever-changing environments.

[0161] In conjunction with the technical solutions of the above embodiments, an embodiment of this application also provides another performance measurement method for a satellite-ground collaborative fine-tuning system. In this method, step S33, "the ground station aggregates multiple received fine-tuning parameters through an aggregation module to obtain the global model parameters for this round," specifically includes steps S33-1 to S33-2:

[0162] Step S33-1: When the fine-tuning parameter transmission of any satellite among the selected satellites participating in parameter aggregation fails, the fine-tuning parameters of multiple satellites that successfully transmitted and the fine-tuning parameters of the target satellite that failed to transmit in the previous round are aggregated to obtain the global model parameters for this round.

[0163] In this embodiment, due to the instability of the link in the satellite-ground collaborative fine-tuning system, the transmission of fine-tuning parameters for some satellites may fail. To ensure the accuracy and completeness of the global model parameters, the aggregation module of the ground station needs to process the received fine-tuning parameters. Specifically, when the transmission of fine-tuning parameters for any of the selected satellites participating in parameter aggregation fails, the aggregation module aggregates the fine-tuning parameters of the multiple satellites that successfully transmitted their parameters from the selected satellites participating in parameter aggregation, along with the fine-tuning parameters of the target satellite that failed to transmit in the previous round, to obtain the global model parameters for this round.

[0164] Specifically, in actual satellite-to-ground communication, due to issues such as high packet loss rates, long latency, or intermittent connections, the fine-tuning parameters of some satellites may fail to be transmitted to the ground station. To address this, the aggregation module retains the successfully transmitted fine-tuning parameters from the previous round and aggregates them together with the successfully transmitted fine-tuning parameters from the current round. The aggregation module performs a weighted average or other aggregation algorithm on the received fine-tuning parameters to generate the global model parameters for this round. This approach ensures that even with partial data loss, the global model parameters still reflect the updates from all participating satellites, thereby maintaining the model's stability and convergence.

[0165] For example, suppose that in a certain training round, three satellites (satellites A, B, and C) participate in parameter aggregation. The fine-tuning parameters of satellites A and B are successfully transmitted to the ground station, but the fine-tuning parameters of satellite C fail to be transmitted. At this time, the aggregation module will aggregate the fine-tuning parameters of satellites A and B together with the fine-tuning parameters of satellite C from the previous round to generate the global model parameters for this round.

[0166] Step S33-2: When the fine-tuning parameters of all selected satellites participating in parameter aggregation are successfully transmitted, the fine-tuning parameters of all selected satellites participating in parameter aggregation are aggregated to obtain the global model parameters for this round.

[0167] In this embodiment, when the fine-tuning parameters of all satellites selected for parameter aggregation are successfully transmitted, the aggregation module aggregates the fine-tuning parameters of all satellites selected for parameter aggregation to obtain the global model parameters for this round.

[0168] Ideally, the fine-tuning parameters of all satellites participating in parameter aggregation are successfully transmitted to the ground station. At this point, the aggregation module directly aggregates these fine-tuning parameters to generate the global model parameters for this round. The aggregation module can use weighted averaging or other aggregation algorithms to weight the fine-tuning parameters based on factors such as the data quality and computing power of each satellite, generating the global model parameters for this round. This approach ensures that the global model parameters accurately reflect the updates of all participating satellites, thereby improving the model's convergence speed and accuracy.

[0169] For example, suppose that in a certain training round, three satellites (satellites A, B, and C) participate in parameter aggregation, and the fine-tuning parameters of all satellites are successfully transmitted to the ground station. At this time, the aggregation module will aggregate the fine-tuning parameters of satellites A, B, and C to generate the global model parameters for this round.

[0170] The technical solution described above enables the system to maintain the continuity and stability of global model parameters even in the event of transmission failure. This is achieved by retaining the fine-tuning parameters from the previous round and aggregating them with the parameters successfully transmitted in the current round. This mechanism allows the system to operate normally under unstable link conditions, reducing the impact of data loss on model training. When all parameters are successfully transmitted, all fine-tuning parameters are directly aggregated to generate the global model parameters for the current round. This ensures that the global model parameters accurately reflect the updates from all participating satellites, thereby improving the model's convergence speed and accuracy. The system effectively handles both partial and complete transmission failures. This flexibility allows the system to adapt to different link conditions and application scenarios, exhibiting high adaptability.

[0171] As a preferred technical solution, to address the situation where fine-tuning parameter transmission fails, network coding technology can be used to encode multiple fine-tuning parameters, ensuring that even if these parameters are lost during transmission, they can still be recovered through other paths. This technology can be applied in satellite-to-ground link environments to mitigate packet loss and reduce transmission latency. By dynamically selecting the optimal coding strategy, the robustness and efficiency of data transmission can be improved to a certain extent.

[0172] As a preferred technical solution, model compression technology can also be used for transmitting fine-tuning parameters. By compressing the fine-tuning parameters of the spaceborne model (such as through low-rank decomposition, pruning, and quantization), the amount of data transmitted each time is reduced, thereby alleviating the network transmission burden, especially when the bandwidth of the space-to-ground link is limited. This solution can be used in conjunction with UDP or TCP protocols to further improve transmission efficiency and reduce the impact of packet loss on model training.

[0173] As a preferred technical solution, when the target protocol is TCP, the high packet loss rate and long latency issues can be addressed by optimizing the TCP congestion control algorithm. For example, combining TCP congestion control algorithms (such as BBR and CUBIC) with a fast recovery mechanism can improve the performance of the TCP protocol to some extent when the satellite-to-ground link is unstable, thereby enhancing the stability of the TCP protocol.

[0174] Figure 3 This is a schematic diagram of the framework of a performance measurement device for a satellite-ground coordinated fine-tuning system according to an embodiment of this application, with reference to... Figure 3 Based on the same inventive concept, another embodiment of this application also provides a performance measurement device for a satellite-ground coordinated fine-tuning system, the device comprising:

[0175] Module 11 is used to construct a satellite-ground coordinated fine-tuning system, which includes multiple satellites and a ground station. The satellites are configured with a local training module to perform local training of the onboard model, generate fine-tuning parameters, and send them to the ground station. The ground station is configured with a satellite selection module and an aggregation module. The satellite selection module is used to select satellites to participate in parameter aggregation, and the aggregation module is used to aggregate the fine-tuning parameters of the selected satellites to generate global model parameters.

[0176] The first measurement module 12 is used to measure the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions based on TCP and UDP protocols respectively, and obtain a first measurement result. The first measurement result is used to represent the target protocol adapted to each unstable link condition.

[0177] The second measurement module 13 is used to measure the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions based on the first measurement result and various priority transmission strategies, respectively, to obtain the second measurement result; the priority transmission strategy is used to indicate the order of transmitting the fine-tuning parameters, and the second measurement result is used to indicate the target priority transmission strategy adapted to each unstable link condition under the target protocol adapted to each unstable link condition.

[0178] Optionally, the first measurement module 12 includes:

[0179] The first simulation unit is used to simulate various unstable link conditions using network simulation tools, including packet loss rate, latency, and network fluctuations.

[0180] The first initialization unit is used to initialize the onboard model and training rounds;

[0181] The first communication link establishment unit is used to establish a communication link between the satellite and the ground station using TCP and UDP protocols respectively under different unstable link conditions.

[0182] The first measurement unit is used to train the spaceborne model and record the convergence status of the training. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold.

[0183] The first recording module is used to determine the target protocol for each unstable link condition from the TCP and UDP protocols based on the convergence of each training round.

[0184] Optionally, the second measurement module 13 includes:

[0185] The second simulation unit is used to simulate various unstable link conditions using network simulation tools, including packet loss rate, latency, and network fluctuations.

[0186] The second initialization unit is used to initialize the onboard model and training rounds;

[0187] The second communication link establishment unit is used to establish a communication link between the satellite and the ground station under each unstable link condition, according to the target protocol adapted to that unstable link condition.

[0188] The second measurement unit is used to train the spaceborne model based on multiple priority transmission strategies and record the convergence status of the training. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold.

[0189] The second recording module is used to select the priority transmission strategy with the highest transmission efficiency among the various priority transmission strategies as the target priority transmission strategy for adapting to the unstable link conditions, based on the convergence of the training.

[0190] Optionally, the device further includes:

[0191] The third measurement module is used to train the onboard model based on the UDP protocol when the target protocol is UDP, and record the convergence status of the training. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold.

[0192] The judgment module is used to determine whether the satellite-ground collaborative fine-tuning system should adopt a retransmission mechanism in actual application based on whether the convergence of the training has reached the expected target convergence.

[0193] The first determining module is used to determine that the retransmission mechanism will not be used on the satellite-ground collaborative fine-tuning system when the convergence of the training reaches the expected first target convergence.

[0194] The second determining module is used to determine the retransmission mechanism to be used on the satellite-ground collaborative fine-tuning system when the convergence of the training reaches the expected second target convergence.

[0195] Wherein, the convergence of the first objective is better than that of the second objective; the retransmission mechanism means that after the satellite transmits the fine-tuning parameters of the selected satellites participating in parameter aggregation to the ground station, and when the communication resources of the communication link between the satellite and the ground station are not exhausted, the fine-tuning parameters of the satellites participating in parameter aggregation that were not successfully transmitted will be retransmitted.

[0196] Optionally, the device further includes:

[0197] The satellite selection module is used to select multiple satellites as target satellites from multiple satellites within the connection window according to preset rules, through the satellite selection module of the ground station under the current unstable link conditions; the preset rules are set based on the current status of each satellite, the quality and correlation of the local data collected by the satellites.

[0198] The communication module is used by the multiple target satellites to establish a communication link with the ground station according to the target protocol adapted to the current unstable link conditions, and to transmit the fine-tuning parameters corresponding to each of the multiple target satellites to the ground station according to the target priority transmission strategy adapted to the current unstable link conditions. The fine-tuning parameters are obtained by the target satellites through a local training module based on local data to train the onboard model deployed locally on the target satellites.

[0199] The aggregation module, used by the ground station, aggregates multiple received fine-tuning parameters to obtain the global model parameters for this round.

[0200] The backhaul module, used by the ground station, transmits the global model parameters to the multiple satellites through a communication link established between the ground station and the target protocol adapted to the current unstable link conditions. This allows the multiple satellites to update the current model parameters of their locally deployed onboard models based on the global model parameters. The satellite selection module then selects new target satellites from the multiple satellites to participate in the next round of local training until the onboard model training is completed.

[0201] Specifically, when the spaceborne model is used to implement disaster navigation, the local data is disaster information data collected by the satellite; when the spaceborne model is used to implement Earth observation, the local data is Earth observation data collected by the satellite; and when the spaceborne model is used to implement climate monitoring, the local data is climate monitoring data collected by the satellite.

[0202] Optionally, the aggregation module includes:

[0203] The first aggregation unit is used to aggregate the fine-tuning parameters of multiple satellites that successfully transmitted the fine-tuning parameters and the fine-tuning parameters of the target satellite that failed to transmit in the previous round when the fine-tuning parameters of any satellite among the selected satellites participating in parameter aggregation fail to transmit, in order to obtain the global model parameters for this round.

[0204] The second aggregation unit is used to aggregate the fine-tuning parameters of all selected satellites participating in parameter aggregation when the fine-tuning parameters of all selected satellites participating in parameter aggregation have been successfully transmitted, so as to obtain the global model parameters for this round.

[0205] Based on the same inventive concept, another embodiment of this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the performance measurement method of the satellite-ground collaborative fine-tuning system as described in any of the above embodiments.

[0206] Based on the same inventive concept, another embodiment of this application provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the performance measurement method of the satellite-ground collaborative fine-tuning system as described in any of the above embodiments.

[0207] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Those skilled in the art should understand that the embodiments of this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. The embodiments of this application may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or terminal device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element. The above provides a detailed description of the performance measurement method, apparatus, equipment, and medium for a satellite-ground coordinated fine-tuning system provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. For those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A performance measurement method for a satellite-ground coordinated fine-tuning system, characterized in that, The method includes: A satellite-ground coordinated fine-tuning system is constructed, comprising multiple satellites and a ground station. Each satellite is equipped with a local training module for performing local training of the onboard model, generating fine-tuning parameters, and sending them to the ground station. The ground station is equipped with a satellite selection module and an aggregation module. The satellite selection module is used to select satellites to participate in parameter aggregation, and the aggregation module is used to aggregate the fine-tuning parameters of the selected satellites to generate global model parameters. Based on TCP and UDP protocols respectively, the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions is measured to obtain a first measurement result. The first measurement result is used to represent the target protocol adapted to each unstable link condition. Based on the first measurement result, the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions is measured according to multiple priority transmission strategies, and a second measurement result is obtained; the priority transmission strategy is used to indicate the order of transmitting the fine-tuning parameters, and the second measurement result is used to represent the target priority transmission strategy adapted to each unstable link condition under the target protocol. The convergence of the onboard model of the satellite-to-ground coordination fine-tuning system under various unstable link conditions is measured based on TCP and UDP protocols respectively, yielding the first measurement result, including: Network simulation tools are used to simulate various unstable link conditions, including packet loss rate, latency, and network fluctuations. Initialize the onboard model and training rounds; Under different unstable link conditions, communication links between the satellite and the ground station are established using TCP and UDP protocols respectively; The spaceborne model is trained, and the convergence status of the training is recorded. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold. Based on the convergence of each training round, the target protocol for each unstable link condition is determined from the TCP and UDP protocols. Based on the first measurement result, and according to various priority transmission strategies, the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions is measured to obtain the second measurement result, including: Network simulation tools are used to simulate various unstable link conditions, including packet loss rate, latency, and network fluctuations. Initialize the onboard model and training rounds; Under each unstable link condition, a communication link between the satellite and the ground station is established according to the target protocol adapted to that unstable link condition. The spaceborne model is trained based on various priority transmission strategies, and the convergence status of the training is recorded. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold. Based on the convergence of the training, the priority transmission strategy with the highest transmission efficiency among the various priority transmission strategies is selected as the target priority transmission strategy for adapting to this unstable link condition.

2. The performance measurement method for the satellite-ground coordinated fine-tuning system according to claim 1, characterized in that, The multiple priority transmission strategies include: no priority transmission strategy, gradient assignment priority transmission strategy, and dynamic priority transmission strategy; The no-priority transmission strategy means that the fine-tuning parameters of each satellite after local training are transmitted in sequence. The gradient assignment priority transmission strategy means that the transmission priority of the fine-tuning parameters after local training of each satellite is determined according to the magnitude of the gradient of the fine-tuning parameters. The dynamic priority transmission strategy means that the fine-tuning parameters of each satellite after local training are dynamically determined according to the contribution of each fine-tuning parameter to the convergence speed in each round of training.

3. The performance measurement method for the satellite-ground coordinated fine-tuning system according to claim 1, characterized in that, The method further includes: When the target protocol is UDP, the onboard model is trained based on UDP, and the convergence status of the training is recorded. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold. Based on whether the convergence of the training reaches the expected convergence, it can be determined whether the satellite-ground collaborative fine-tuning system should adopt a retransmission mechanism in actual application. When the training convergence reaches the expected first target convergence, it is determined that the retransmission mechanism will not be used on the satellite-ground collaborative fine-tuning system. When the training convergence reaches the expected second target convergence, it is determined that the retransmission mechanism will be adopted on the satellite-ground collaborative fine-tuning system. Wherein, the convergence of the first objective is better than that of the second objective; the retransmission mechanism means that after the satellite transmits the fine-tuning parameters of the selected satellites participating in parameter aggregation to the ground station, and when the communication resources of the communication link between the satellite and the ground station are not exhausted, the fine-tuning parameters of the satellites participating in parameter aggregation that were not successfully transmitted will be retransmitted.

4. The performance measurement method for the satellite-ground coordinated fine-tuning system according to any one of claims 1-3, characterized in that, The method further includes: Under the current unstable link conditions, the satellite selection module of the ground station selects multiple satellites as target satellites from multiple satellites within the connection window according to preset rules; the preset rules are set based on the current status of each satellite, the quality and correlation of the local data collected by the satellites. The multiple target satellites establish communication links with the ground station through a communication module, according to the target protocol adapted to the current unstable link conditions. According to the target priority transmission strategy adapted to the current unstable link conditions, the fine-tuning parameters corresponding to each of the multiple target satellites are transmitted to the ground station. The fine-tuning parameters are obtained by the target satellites through a local training module based on local data to train the onboard model deployed locally on the target satellites. The ground station uses an aggregation module to aggregate multiple received fine-tuning parameters to obtain the global model parameters for this round. The ground station transmits the global model parameters to the multiple satellites through a communication link established between the ground station and the target protocol adapted to the current unstable link conditions. This enables the multiple satellites to update the current model parameters of their locally deployed onboard models based on the global model parameters. The satellite selection module then selects new target satellites from the multiple satellites to participate in the next round of local training until the onboard model training is completed. Specifically, when the spaceborne model is used to implement disaster navigation, the local data is disaster information data collected by the satellite; when the spaceborne model is used to implement Earth observation, the local data is Earth observation data collected by the satellite; and when the spaceborne model is used to implement climate monitoring, the local data is climate monitoring data collected by the satellite.

5. The performance measurement method for the satellite-ground coordinated fine-tuning system according to claim 4, characterized in that, The ground station, through an aggregation module, aggregates multiple received fine-tuning parameters to obtain the global model parameters for this round, including: When the fine-tuning parameters of any satellite among the selected satellites participating in parameter aggregation fail to be transmitted, the fine-tuning parameters of multiple satellites that successfully transmitted and the fine-tuning parameters of the target satellite that failed to transmit in the previous round are aggregated to obtain the global model parameters for this round. When the fine-tuning parameters of all selected satellites participating in parameter aggregation are successfully transmitted, the fine-tuning parameters of all selected satellites participating in parameter aggregation are aggregated to obtain the global model parameters for this round.

6. A performance measurement device for a satellite-ground coordinated fine-tuning system, characterized in that, The device includes: A construction module is used to build a satellite-ground coordinated fine-tuning system, which includes multiple satellites and a ground station. The satellites are configured with a local training module to perform local training of the onboard model, generate fine-tuning parameters, and send them to the ground station. The ground station is configured with a satellite selection module and an aggregation module. The satellite selection module is used to select satellites to participate in parameter aggregation, and the aggregation module is used to aggregate the fine-tuning parameters of the selected satellites to generate global model parameters. The first measurement module is used to measure the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions based on TCP and UDP protocols respectively, and obtain the first measurement result. The first measurement result is used to represent the target protocol adapted to each unstable link condition. The second measurement module is used to measure the convergence of the onboard model of the satellite-ground coordination fine-tuning system under various unstable link conditions based on the first measurement result and various priority transmission strategies, respectively, to obtain the second measurement result; the priority transmission strategy is used to indicate the order of transmitting the fine-tuning parameters, and the second measurement result is used to represent the target priority transmission strategy adapted to each unstable link condition under the target protocol adapted to each unstable link condition. The first measurement module includes: The first simulation unit is used to simulate various unstable link conditions using network simulation tools, including packet loss rate, latency, and network fluctuations. The first initialization unit is used to initialize the onboard model and training rounds; The first communication link establishment unit is used to establish a communication link between the satellite and the ground station using TCP and UDP protocols respectively under different unstable link conditions. The first measurement unit is used to train the spaceborne model and record the convergence status of the training. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold. The first recording module is used to determine the target protocol for each unstable link condition from the TCP and UDP protocols based on the convergence of each training round. The second measurement module includes: The second simulation unit is used to simulate various unstable link conditions using network simulation tools, including packet loss rate, latency, and network fluctuations. The second initialization unit is used to initialize the onboard model and training rounds; The second communication link establishment unit is used to establish a communication link between the satellite and the ground station under each unstable link condition, according to the target protocol adapted to that unstable link condition. The second measurement unit is used to train the spaceborne model based on multiple priority transmission strategies and record the convergence status of the training. The convergence status includes: training time, convergence time, loss value, and accuracy. The convergence time is the time required for the training to reach a preset error threshold. The convergence time is determined based on the loss value of each training round, the current training round, the training round, and the error threshold. The second recording module is used to select the priority transmission strategy with the highest transmission efficiency among the various priority transmission strategies as the target priority transmission strategy for adapting to the unstable link conditions, based on the convergence of the training.

7. An electronic device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the performance measurement method for the satellite-ground coordinated fine-tuning system as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the performance measurement method of the satellite-ground coordinated fine-tuning system as described in any one of claims 1-5.

Images