A single-satellite system based on decentralized federated learning and a model on-orbit cooperative training method

Abstract

This invention discloses a single-satellite system and an on-orbit collaborative training method for models based on decentralized federated learning. This architecture aims to resolve the conflict between data privacy protection and value circulation faced by a single satellite carrying multi-user tasks, achieved by constructing a logically isolated distributed peer-to-peer network on a single physical satellite computing platform. The system virtualizes onboard computing and storage resources into multiple independent logical nodes. Each node represents a different data owner or algorithm model holder, possessing strictly isolated private storage areas and private computing power quotas, forming a physically unified but logically independent intra-satellite data island environment. Interconnection between logical neighbors can be achieved through secure intra-satellite channels without the need for a central aggregation unit. This invention provides a reliable technical paradigm for the secure circulation of satellite data in sensitive scenarios such as commercial sharing, and is of great significance for promoting the asset-based management and intelligent application of satellite data.

Description

Technical Field

[0001] This invention relates to a single satellite system and a method for on-orbit collaborative training of models based on decentralized federated learning, belonging to the field of satellite system design. Background Technology

[0002] With the rapid development of satellite intelligence and satellite internet, satellites are playing an increasingly prominent role in fields such as communication, remote sensing, meteorology, and navigation. The real-time, multi-source data they collect provides valuable resources for scientific research and commercial applications. However, satellite data is often widely distributed and involves sensitive information, posing serious challenges to traditional data processing and sharing methods, especially given the exacerbating problem of data silos. This leads to difficulties in data integration and limited intelligent analysis capabilities. The satellite data silo problem stems from several factors: First, data is scattered across multiple payloads / computing nodes within the satellite, lacking efficient distributed collaboration mechanisms, hindering data flow; second, sensitive data involving security and commerce restricts direct exchange, and traditional centralized processing methods struggle to guarantee privacy; finally, inconsistent standards among different satellite payloads, with significant differences in data formats, storage, and processing procedures, further increase the complexity of collaboration.

[0003] Federated learning, as a distributed machine learning paradigm, offers an innovative approach to solving the aforementioned problems. In particular, decentralized federated learning architectures, by removing a central server and employing peer-to-peer networks for collaborative training, exchange only encrypted model parameters rather than raw data, effectively protecting data privacy and avoiding the risk of leakage. In satellite scenarios, decentralized architectures utilize direct inter-satellite communication and consensus mechanisms, enabling each satellite node to independently complete local model training and update the global model through distributed aggregation, thereby breaking down data silos and promoting efficient utilization of data resources. The increased onboard computing power brought about by satellite intelligence provides a foundation for decentralized training, supporting real-time on-orbit data processing and model optimization, while satellite internet technology enhances inter-node interconnectivity and reduces communication latency and costs. However, general decentralized federated learning mechanisms cannot be directly transferred to single-satellite onboard scenarios. The main reasons are as follows: First, the computing power, storage, energy, and intra-satellite bus bandwidth of a single satellite platform are all strictly constrained. If multiple logical nodes train and broadcast parameters concurrently without restraint in a ground-based distributed learning manner, it is easy to cause computing power contention, cache conflicts, and intra-satellite communication congestion. Second, different logical nodes within a single satellite usually correspond to different payloads, tasks, data ownership, or model uses. The data held by each node have significant differences in imaging mechanisms, observation time phases, regional coverage, category distribution, and annotation quality. Non-independent and identically distributed data can easily cause local model drift and model oscillation after aggregation. Third, the decentralized architecture lacks a central server for unified verification and scheduling. If the adjacency relationships, parameter versions, update integrity, and abnormal offsets between logical nodes are not constrained, expired, missing, or abnormal model updates may continue to propagate in the peer-to-peer network, affecting model convergence consistency and on-orbit operational stability. Therefore, single-satellite on-board federated learning cannot be achieved simply by deploying a general decentralized algorithm. Instead, it requires satellite-based modifications to node organization, training budget, topology communication, parameter verification, and aggregation mechanisms, taking into account the characteristics of limited on-board resources, logical isolation, multi-task sharing, and secure parameter interaction.

[0004] Decentralized federated learning reduces communication overhead through sparse topology and local interactions, and enhances security using privacy amplification techniques, making it suitable for dynamic and unreliable satellite network environments. With the development of low-Earth orbit satellites and large-scale satellite networks, decentralized federated learning has broad application prospects in the satellite field. It not only improves data utilization efficiency and system robustness, but also avoids single points of failure through distributed consensus, promoting the intelligentization of satellite data and the collaborative development of global networks. In the future, this architecture will play a core role in satellite data sharing, real-time analysis, and secure collaboration. Summary of the Invention

[0005] The purpose of this invention is to address the single-point failure risk, communication bottlenecks, and scalability limitations of existing centralized federated learning architectures in satellite environments by providing a decentralized federated learning method for single-satellite systems and on-orbit collaborative training of models.

[0006] This invention is achieved through the following technical solution: A single-satellite system based on decentralized federated learning constructs a distributed peer-to-peer federated learning network composed of multiple logical nodes on a single satellite computing platform. It eliminates the need for a central aggregation server. Each independent logical node possesses a private storage area and private computing power quota for local dataset management and model training. Each node autonomously generates format-aligned and labeled data on-orbit for local training based on its acquired data. Through a secure intra-satellite communication channel, it interacts only with its adjacent logical nodes based on the intra-satellite logical interconnection topology, exchanging model parameters or parameter updates, and performing consensus aggregation locally. The intra-satellite logical interconnection topology is represented by a connected graph, used to define the adjacency relationships and parameter interaction paths between logical nodes, thereby avoiding full-connection broadcasting while reducing intra-satellite communication overhead.

[0007] In the above technical solution, the private storage area is a private dedicated storage space opened up on the satellite for each node, used to store the dedicated private datasets and local model parameters of each node respectively, ensuring that the data is stored locally on the node and not leaked, and supporting non-independent and identically distributed data distribution.

[0008] Furthermore, the private computing power quota is the private dedicated on-board computing power allocated to each node on the satellite. It is used by each node to perform local stochastic gradient descent or other optimization algorithms using its private dataset to train local models and reduce communication frequency through multiple local update steps.

[0009] Furthermore, local training on each node employs a budget-based training mechanism. Before initiating a round of local training, each node determines its training budget based on its allocated private computing power quota, available storage space, model cache usage, intra-satellite bus load, and current task priority. This training budget constrains the number of local iterations, training duration, gradient cache size, and parameter exchange frequency for that round. Specifically, a resource monitor and a training budget controller are set up within each logical node. Before starting a round of training, the resource monitor collects the following indicators for that logical node: remaining computing power ratio, available storage ratio, model cache usage ratio, intra-satellite bus idle ratio, and task priority, and normalizes these indicators into state variables between 0 and 1. The training budget controller weights these state variables according to preset weights to obtain the budget coefficient for that round, where the model cache usage ratio and intra-satellite bus load are used with negative weights, and the task priority is used with positive weights. Subsequently, the training budget controller determines the maximum number of local iterations, maximum training duration, maximum gradient cache capacity, and minimum parameter exchange interval for that round based on the budget coefficient and the configuration limit of the logical node.

[0010] Furthermore, during local training, a local optimization method adapted to non-independent and identically distributed data is adopted, and a offset constraint relative to the current shared model state is applied to the local update process, thereby suppressing model drift caused by local data bias.

[0011] Furthermore, the logical interconnection topology is constructed based on task relevance, data complementarity, node historical stability, and intra-satellite communication load status. Each node interacts with only a limited number of neighboring nodes based on the pre-constructed logical interconnection topology, without adopting a fully connected broadcast method.

[0012] Furthermore, after a node communicates with a limited number of neighbors, the neighboring nodes first perform validity verification on the received parameters. The verification includes: source node identity verification, model version consistency verification, parameter integrity verification, timestamp validity verification, and abnormal offset detection. Only when the received parameters pass the validity verification will the node enter the local aggregation stage.

[0013] Furthermore, each node independently executes a decentralized aggregation algorithm locally based on its own local update results and valid parameters from neighboring nodes, applying dynamic aggregation weights to the multiple update amounts to form the aggregation model state for the current round, which serves as the initial model state for the next round of local training.

[0014] A method for on-orbit collaborative training of a single-satellite federated learning model, based on the single-satellite system described in any of the preceding methods, includes the following: Each node first uses its private computing power in its local private storage area to complete on-orbit preprocessing and sample consistency processing of the data, thereby achieving on-orbit autonomous production of format-aligned and labeled data. Under the constraint of a pre-determined training budget for this round, each node independently performs local model training based on its own private training samples. After local training is completed, each node interacts with only a limited number of neighboring nodes based on the pre-constructed logical interconnection topology. After receiving the parameters, each neighboring node performs parameter validity verification. If the verification is successful, the node enters the local aggregation stage. Based on its own local update results and the valid parameters from the neighboring nodes, each node independently executes a decentralized aggregation algorithm with dynamic weights to form the aggregated model state of the current round, which serves as the initial model state for the next round of local training. As the number of iterations increases, training ends when the preset convergence condition is met or the maximum number of iterations is reached. Each node then solidifies the finally converged model parameters, forming a consistent converged model copy, and saves it in its own private storage area.

[0015] Compared with the existing centralized architecture, this invention eliminates the risk of single point of failure through decentralized design, improves the robustness and scalability of the system in the dynamic satellite network environment; significantly reduces satellite-to-ground communication latency and bandwidth pressure through local parameter exchange and consensus aggregation, and further ensures data security by utilizing privacy enhancement technology. Compared to other existing satellite systems employing decentralized federated learning, this invention does not simply treat each satellite as a physical federated node. Instead, it constructs multiple logically isolated nodes within a single satellite computing platform. This allows different payloads, tasks, or data owners on the same satellite to collaborate on models without sharing raw data or establishing a central aggregation unit. Furthermore, this invention addresses the resource contention issue of multiple tasks on a single satellite, which is not adequately considered in existing constellation-level decentralized solutions, through private storage areas and private computing power quotas. It also addresses the problem of uncontrolled concurrent training caused by limited onboard computing power, cache, and intra-satellite bus through budgeted local training. The invention resolves the intra-satellite communication overhead caused by fully connected broadcasts through intra-satellite logical interconnection topology. It addresses the problem of abnormal updates, inconsistent versions, or the continued propagation of outdated parameters in a decentralized architecture through parameter validity verification. Finally, it mitigates model drift and aggregation oscillations caused by non-independent and identically distributed data through dynamic weight aggregation. Therefore, this invention solves the problems of logical isolation of multiple tenants within a single satellite, resource-constrained training, reliable exchange of intra-satellite parameters, and consistent convergence, which are often overlooked in existing decentralized satellite federated learning schemes that focus on cross-satellite collaboration.

[0016] The advantages of this invention compared to the prior art are as follows: This invention innovatively proposes a decentralized federated learning satellite architecture. This architecture achieves federated learning capabilities based on a distributed peer-to-peer network, eliminating the need for a central server. Each satellite node directly interacts and aggregates model parameters through direct inter-satellite communication and a distributed consensus mechanism. This not only effectively eliminates the risk of single points of failure and significantly improves the system's robustness and reliability in dynamic satellite network environments, but also, by adhering to the principle of sharing only privacy-preserving model parameters and not raw data, breaks down data silos, achieves cross-node data value sharing, and further strengthens data privacy and security, reducing the potential risk of data leakage. This invention proposes a decentralized federated learning method for on-orbit collaborative training of satellite models. By utilizing format-aligned labeled data generated autonomously on the satellite and combining it with direct collaborative training between nodes, it can provide richer and more effective training datasets for on-orbit AI models. This method particularly focuses on addressing the inherent heterogeneity challenges of satellite data. Through a distributed collaborative mechanism, it can better utilize the diverse data collected by different satellites, thereby effectively overcoming the limitations imposed on model performance by data scarcity or uneven distribution, significantly improving the accuracy and practicality of satellite intelligent models, and accelerating the satellite intelligence process.

[0017] The decentralized architecture and parameter-sharing model adopted in this invention provide a more solid technical foundation for the controlled circulation and application of satellite data. The distributed decision-making mechanism reduces reliance on a single trusted entity and enhances the system's resistance to attacks and its self-recovery capabilities. Simultaneously, this architecture supports the construction of new satellite business models based on model parameter collaboration, such as establishing a distributed parameter market. This enables the value transfer of satellite data without moving the original data, opening up new paths for promoting the assetization and commercial application of satellite data, and possessing significant engineering application value and commercial prospects. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the single-satellite architecture based on decentralized federated learning in this invention; Figure 2 This is a schematic diagram of the on-orbit collaborative training method for the decentralized federated learning model of the present invention. Detailed Implementation

[0019] The technical solution of this invention will be further described in detail below with reference to specific embodiments. The satellite described below already possesses high-speed networked communication capabilities between space and ground and sufficient computing power to support decentralized federated learning. The satellite architecture provided by this invention aims to solve the contradiction between data privacy protection and value circulation faced by a single satellite when carrying multi-user (such as multi-departmental collaboration) tasks. This is achieved by constructing a logically isolated distributed peer-to-peer network on a single physical satellite computing platform. The system virtualizes onboard computing and storage resources into multiple independent logical nodes (Parties). Each node represents a different data owner or algorithm model holder, possessing strictly isolated private storage areas and private computing power quotas, forming a physically unified but logically independent intra-satellite data island environment. Interconnection between logical neighbors can be achieved through secure intra-satellite channels without the need for a central aggregation unit. This invention breaks through the limitations of traditional centralized data processing models through an innovative mechanism of intra-satellite logical isolation and secure parameter interaction. It not only effectively avoids the risk of raw data leakage in multi-tenant scenarios and ensures the data security of all participants, but also utilizes on-board edge computing capabilities to achieve real-time fusion and value mining of heterogeneous data, significantly improving the generalization ability and robustness of single-star intelligent models in complex scenarios.

[0020] To address the challenge of directly transferring general decentralized federated learning to a single-satellite scenario, this embodiment employs the following satellite-based processing methods: A single satellite platform is divided into multiple logically isolated nodes through resource virtualization, with private storage areas and private computing power quotas defining the data boundaries and training resources of each node; In-orbit preprocessing and format alignment marking modules convert data from different payloads and task sources into training samples that meet the unified model input requirements; A budget-based training mechanism determines the training budget for each round based on computing power quotas, storage capacity, model cache usage, intra-satellite bus load, and task priority, preventing resource congestion caused by concurrent training of multiple logical nodes; A pre-constructed intra-satellite logical interconnection topology defines the adjacency relationships of each node, ensuring that nodes exchange parameters only with a limited number of neighbors rather than through full-connection broadcasting; Invalid or abnormal parameters are blocked from entering the aggregation process through source identity, model version, parameter integrity, timestamps, and abnormal offset verification; A dynamic weight aggregation mechanism adjusts the impact of neighbor updates based on sample quality, task relevance, parameter freshness, and historical stability. Therefore, this embodiment can reduce intra-satellite communication and caching overhead without sharing the original observation data, suppress model drift caused by non-independent and identically distributed data, improve the robustness of the centerless aggregation process to abnormal updates and resource fluctuations, and enable multiple logical data silos within a single satellite platform to form a consistent and convergent onboard intelligent model.

[0021] This embodiment constructs a logically isolated decentralized federated learning network on a single physical satellite platform. Specifically, the satellite computing platform uses resource virtualization to divide the task units corresponding to different members or organizations into multiple independent logical nodes. Each logical node is configured with a private storage area, private computing power quota, local training environment, and parameter exchange interface, thereby forming a physically unified but logically isolated intra-satellite peer-to-peer collaborative structure. Further, these logical nodes can correspond to maritime monitoring task nodes, meteorological observation task nodes, commercial remote sensing task nodes, or other on-board intelligent processing task nodes. Through this setup, the data, models, and training processes of different members all run independently within their respective private domains, achieving decentralized model collaboration without relying on a central aggregation unit.

[0022] Furthermore, the raw observation data collected by the onboard sensors is distributed to the private storage area of ​​the corresponding logical node at the source of data generation, based on task scheduling strategies, member permissions, or model usage, thereby avoiding cross-access of raw data between different members. Considering the differences in the service types, observation payloads, time conditions, and target scenarios served by different logical nodes within a single physical satellite platform, the data held by each node exhibits significant non-independent and identically distributed characteristics. For example, some logical nodes mainly process optical remote sensing data acquired during the day, some mainly process SAR data acquired at night or under complex weather conditions, and some logical nodes may only cover densely populated port areas or sparsely populated open ocean areas. Based on this, this embodiment does not directly use the raw data for joint training, but requires each logical node to first complete on-orbit preprocessing and sample consistency processing within its local private domain.

[0023] After the data enters the private storage area, the onboard preprocessing pipeline is first triggered. For optical remote sensing data, geometric correction, radiometric correction, invalid region removal, and scale normalization are preferably performed; for SAR remote sensing data, speckle suppression, geocoding, amplitude normalization, and tensor structure transformation are preferably performed. Subsequently, each logical node further performs format alignment processing on the preprocessed data to ensure that data from different payload sources meet the unified model training requirements in terms of input size, feature structure, and annotation format. Furthermore, to achieve on-orbit autonomous production of format-aligned labeled data, this embodiment deploys an automatic annotation module within each logical node. Taking ship detection as an example, real-time messages received by the onboard AIS can be used as weak supervision signals to perform temporal alignment, attitude correction, and spatial matching between AIS location information and remote sensing images, automatically generating location labels or bounding box labels for ship targets; for areas with missing AIS signals, insufficient matching, or target occlusion, a pseudo-label generation module is further invoked to generate supplementary labels based on the current detection model output results, candidate box confidence, and spatiotemporal continuity constraints. Using the above method, each logical node can independently construct private training samples with uniform format, aligned features, and suitable for subsequent federated training without relying on manual annotation on the ground.

[0024] In this embodiment, the local training does not employ an unconstrained continuous training approach, but rather a budget-based training mechanism oriented towards the on-board shared platform. Before initiating a round of local training, each logical node determines its training budget based on its allocated private computing power quota, available storage space, model cache usage, onboard bus load, and current task priority. This training budget constrains the number of local iterations, training duration, gradient cache size, and parameter exchange frequency for that round. This avoids computing power contention, cache conflicts, and communication congestion that can occur when multiple logical nodes run simultaneously on the same physical satellite platform. Subsequently, each logical node independently executes local model training based on its own private training samples. Considering the highly heterogeneous data distribution in a single-satellite, multi-member scenario, this embodiment preferably adopts a local optimization method adapted to non-independent, identically distributed data, applying offset constraints relative to the current shared model state to the local update process, thereby suppressing model drift caused by local data bias. As a specific implementation, each logical node is equipped with a resource monitor and a training budget controller. Before the start of the current training round, the resource monitor collects the remaining computing power ratio, available storage ratio, model cache usage ratio, intra-satellite bus idle ratio, and task priority of the logical node, and normalizes these indicators into state variables between 0 and 1. The training budget controller weights these state variables according to preset weights to obtain the budget coefficient for this round, where the model cache usage ratio and intra-satellite bus load are used with negative weights, and task priority is used with positive weights. Subsequently, the training budget controller determines the maximum number of local iterations, maximum training duration, maximum gradient cache capacity, and minimum parameter exchange interval for this round based on the budget coefficient and the configuration limit of the logical node. For example, when the budget coefficient is lower than a first threshold, the logical node only performs a small batch of updates and extends the parameter exchange interval; when the budget coefficient is higher than a second threshold, the logical node is allowed to perform more local iterations and shorten the parameter exchange interval; when the intra-satellite bus load or model cache usage exceeds a preset safety threshold, even if the maximum number of iterations has not been reached, the current training round is stopped early, and the completed parameter updates are written to the local cache.

[0025] As one implementation of local optimization for non-independent and identically distributed data, each logical node adds a proximal constraint term relative to the current shared model state to its local loss function. This means that the greater the deviation of the local model parameters from the previous round's aggregated model state, the greater the corresponding penalty. Alternatively, after each round of local updates, norm pruning is performed on the parameter update amount. When the update amount exceeds a preset offset threshold, it is scaled proportionally to within the threshold range. Through these proximal constraint or update pruning methods, it is possible to limit excessive model drift caused by local sample distribution bias while maintaining the ability of each node to learn task features using local private samples.

[0026] After local training is complete, each logical node outputs the updated parameters or gradient representation of the local model for that round, without outputting the original data or intermediate samples. If a logical node detects during training that local resource usage has reached a threshold, intra-satellite bus waiting time has exceeded a threshold, or the current round of scheduling time slice is about to end, it terminates the current round of training early and saves the currently completed update results in the local cache for subsequent parameter exchange. This makes the decentralized federated training process more adaptable to the actual operating environment of multi-task sharing and resource constraints within a single physical satellite platform.

[0027] During the parameter exchange phase, each logical node interacts with only a limited number of neighboring nodes based on a pre-constructed logical interconnection topology, rather than using a fully connected broadcast approach. The logical interconnection topology is preferably constructed based on task relevance, data complementarity, node historical stability, and intra-satellite communication load status, thereby ensuring that model information can be gradually propagated throughout the network while reducing the link burden of a single round of parameter exchange. Furthermore, before sending local parameters, each logical node performs privacy enhancement processing on its local updates to ensure member data security and model security; this privacy enhancement processing can employ differential privacy noise addition, homomorphic encryption encapsulation, or other parameter protection methods. Afterward, each logical node sends the processed parameters to its logical neighboring nodes through an intra-satellite secure channel. To further reduce intra-satellite communication overhead, this embodiment can also perform block transmission, sparsity filtering, or priority sorting of the parameters to be sent, prioritizing the transmission of parameters more sensitive to the current model's convergence, while storing the remaining unsent parameters in a local buffer for exchange in the next scheduling cycle.

[0028] Upon receiving parameters, each logical neighbor node does not immediately perform fusion; instead, it first verifies the validity of the parameters. The verification preferably includes: source node identity verification, model version consistency verification, parameter integrity verification, timestamp validity verification, and abnormal offset detection. Parameters that fail verification, are outdated, have incomplete transmission, or exhibit abnormal update magnitudes may be discarded, downgraded, or temporarily deferred from inclusion in the current aggregation round. Only after the received parameters pass validity verification does the local aggregation phase begin. By adding this verification step between parameter reception and fusion, the stability and anomaly resistance of decentralized federated learning in complex onboard environments can be improved.

[0029] During the local aggregation phase, each logical node independently executes a decentralized aggregation algorithm based on its own local update results and valid parameters from neighboring nodes, forming the aggregated model state for the current round. Furthermore, this embodiment does not simply average the update amounts from multiple parties, but rather determines dynamic aggregation weights based on neighboring node sample quality, task relevance, parameter freshness, historical training stability, and version consistency. For neighboring nodes that are more relevant to the current node's task scenario, have more recent parameter updates, and higher historical training stability, the aggregation weight of their corresponding parameters is increased; for neighboring nodes with lagging versions, large fluctuations, or low task relevance, the influence of their corresponding parameters is reduced. Thus, the aggregated model state retains the task characteristics of the current node while gradually absorbing complementary knowledge formed by other logical nodes under different observation scenarios. After aggregation, the aggregated model state serves as the initial model state for the next round of local training, continuing to participate in subsequent local training, neighbor exchange, and local aggregation cycles.

[0030] As the number of iterations increases, the model parameters held by each logical node gradually become consistent. The decentralized federated training process stops when a preset convergence condition is met or the maximum number of iterations is reached. The preferred convergence conditions include: the difference in model parameters is below a preset threshold for several consecutive iterations, the decrease in local loss is below a preset threshold, or the system reaches the maximum number of training iterations. After training, each logical node solidifies the finally converged model parameters, forming a consistent, converged model copy, and stores it in its own private storage area. Thus, without sharing the original observation data, each member can obtain a high-performance spaceborne intelligent model that integrates the observation knowledge of multiple members, thereby completing decentralized on-orbit collaborative training within a single physical satellite platform.

[0031] The above description of the disclosed examples enables those skilled in the art to implement or use the present invention. Various modifications to these examples will be readily apparent to those skilled in the art. The examples defined herein are merely instances of the decentralized federated learning satellite architecture and on-orbit collaborative training method of the present invention; however, simple modifications and substitutions to the algorithm model, Local model training strategy, Global model aggregation method, privacy implementation of model parameter transmission, and privacy implementation of data storage and access should all fall within the protection scope of the corresponding claims of the present invention. Therefore, the present invention is not limited to the example shown herein, but is to be accorded the widest scope consistent with the principles disclosed herein.

Claims

1. A single-satellite system based on decentralized federated learning, characterized in that, A distributed peer-to-peer federated learning network consisting of multiple logical nodes is constructed on a single satellite computing platform. There is no need to set up a central aggregation server. Each independent logical node has a private storage area and a private computing power quota for local dataset management and model training. Each node autonomously generates format-aligned and labeled data on orbit for its acquired data and performs local training. Through the intra-satellite secure communication channel, it only interacts with its neighboring logical nodes according to the intra-satellite logical interconnection topology to exchange model parameters or parameter update amounts, and performs consensus aggregation locally. The intra-satellite logical interconnection topology is represented by a connected graph to limit the adjacency relationship and parameter interaction path between logical nodes, so as to avoid full-connection broadcasting and reduce intra-satellite communication overhead.

2. The single-satellite system based on decentralized federated learning according to claim 1, wherein, The private storage area is a private dedicated storage space allocated on the satellite for each node, used to store the dedicated private datasets and local model parameters of each node, ensuring that the data is stored locally on the node and not leaked, and supporting non-independent and identically distributed data distribution.

3. The single-satellite system based on decentralized federated learning of claim 1, wherein, The private computing power quota is the private dedicated on-board computing power allocated to each node on the satellite. It is used by each node to perform local stochastic gradient descent or other optimization algorithms to train local models using its private dataset, and to reduce communication frequency through multiple local update steps.

4. The single-satellite system based on decentralized federated learning of claim 1, wherein, The local training of each node adopts a budget-based training mechanism. Before starting the local training round, each node first determines the training budget for the current round based on its allocated private computing power quota, available storage space, model cache usage, intra-satellite bus load, and current task priority. The training budget is used to constrain the number of local iterations that can be executed in the current round, the training duration, the gradient cache size, and the parameter exchange frequency.

5. The single-satellite system based on decentralized federated learning according to claim 4, characterized in that, Within each logical node, a resource monitor and a training budget controller are set up. Before the start of this round of training, the resource monitor collects the following indicators of the logical node: remaining computing power ratio, available storage ratio, model cache usage ratio, intra-satellite bus idle ratio, and task priority, and normalizes the above indicators into state variables between 0 and 1. The training budget controller weights the aforementioned state variables according to preset weights to obtain the budget coefficients for this round. The model cache occupancy ratio and intra-satellite bus load are used in the calculation with negative weights, while the task priority is used in the calculation with positive weights. Subsequently, the training budget controller determines the maximum number of local iterations, the maximum training duration, the maximum gradient cache capacity, and the minimum parameter exchange interval for this round based on the budget coefficients and the configuration limit of the logical node.

6. The single-satellite system based on decentralized federated learning of claim 1, wherein, During local training, a local optimization method adapted to non-independent and identically distributed data is adopted. An offset constraint relative to the current shared model state is applied to the local update process, thereby suppressing model drift caused by local data bias.

7. The single-satellite system based on decentralized federated learning of claim 1, wherein, The logical interconnection topology is constructed based on task relevance, data complementarity, node historical stability, and intra-satellite communication load status. Each node interacts with only a limited number of neighboring nodes based on the pre-constructed logical interconnection topology, without using a fully connected broadcast method.

8. The single-satellite system based on decentralized federated learning of claim 1, wherein, After a node communicates with a limited number of neighbors, the neighboring nodes first verify the validity of the received parameters. The verification includes: source node identity verification, model version consistency verification, parameter integrity verification, timestamp validity verification, and abnormal offset detection. Only when the received parameters pass the validity verification will the node enter the local aggregation stage.

9. The single-satellite system based on decentralized federated learning according to claim 1, characterized in that, Each node independently executes a decentralized aggregation algorithm locally based on its own local update results and valid parameters from neighboring nodes. It uses dynamic aggregation weights to form the aggregation model state for the current round, which serves as the initial model state for the next round of local training.

10. A method for on-orbit cooperative training of a single-satellite federated learning model, characterized in that, Based on the single-satellite system as described in any one of claims 1-9, the method includes the following: Each node first uses its private computing power in its local private storage area to complete on-orbit preprocessing and sample consistency processing of the data, thereby achieving on-orbit autonomous production of format-aligned and labeled data. Under the constraint of a pre-determined training budget for this round, each node independently performs local model training based on its own private training samples. After local training is completed, each node interacts with only a limited number of neighboring nodes based on the pre-constructed logical interconnection topology. After receiving the parameters, each neighboring node performs parameter validity verification. If the verification is successful, the node enters the local aggregation stage. Based on its own local update results and the valid parameters from the neighboring nodes, each node independently executes a decentralized aggregation algorithm with dynamic weights to form the aggregated model state of the current round, which serves as the initial model state for the next round of local training. As the number of iterations increases, training ends when the preset convergence condition is met or the maximum number of iterations is reached. Each node then solidifies the finally converged model parameters, forming a consistent converged model copy, and saves it in its own private storage area.

Images