低轨卫星网络的在轨协同模型训练方法、系统及存储介质
By constructing a two-dimensional split training architecture and a pipelined parallel scheduling mechanism in low-Earth orbit satellite networks, and combining multi-objective optimization and global optimization algorithms, the problems of large transmission latency, high energy consumption and slow model convergence in low-Earth orbit satellite networks are solved, and efficient and low-consumption on-orbit collaborative training is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-17
AI Technical Summary
Low-Earth orbit satellite networks suffer from problems such as large transmission delays, high energy consumption, insufficient data backhaul security, slow model convergence, and node energy overload. Existing solutions cannot effectively address these issues.
We construct an on-orbit collaborative model training method adapted to the resource-constrained characteristics of low-Earth orbit satellite networks. By using a two-dimensional split training architecture and a pipelined parallel scheduling mechanism, combined with a multi-objective optimization model and a global optimization algorithm, we determine the optimal model splitting strategy and collaborative node configuration to achieve temporal overlap between on-board computing and inter-satellite transmission.
It significantly reduces computational latency and energy consumption, improves the adaptability and efficiency of model training, ensures model performance, and solves the shortcomings of traditional centralized training and federated learning schemes.
Smart Images
Figure CN122414291A_ABST