面向卫星在轨处理的强化学习任务调度算法

打开文本图片集

中图分类号：TP391 文献标志码：A DOI：10.12305/j.issn.1001-506X.2025.06.20

Abstract：As satellite earth observation enters anera multiple satelites，high resolution，real-time response，and global observation，satelite on-orbit data processg has become one the ma methods to improve the real-time characteristic remote sensg data processg.In scenarios where satelite resources are limited，data transmission lk channels are constraed，and opportunisticobservation tasks are unpredictable， real-time schedulg data processg tasks faces significant challenges.An optimization problem model with thegoal maximizg the system's averagedata processg throughput rate is firstly constructed.Secondly，an onle task schedulg algorithm thatcombes deep reforcement learng（DRL） is proposed.DRL algorithm enablesreal-timecalculation task schedulg strategies，and Lagrangian dual optimizationalgorithm can accurately computes the optimal resource allcation.Fally，simulation experiments are conducted to evaluate the effectiveness and data processg throughput rate the proposed algorithm.Results show that the proposed algorithm can converge and approach the optimal solution，improvg data processng throughput rate by approximately 8% compared to existg algorithms，and demonstratg scalability as the satelite data arrival speed and the number satelite computg nodes crease.Theproposed algorithm can maximize the average data processg throughput rate the system while ensurg the stability and convergence task queue length and average energy consumption a high-dynamic environment.

Keywords：satelite on-orbit processg；task schedulg；resource allocation；deep reforcement learng （DRL）；Lyapunov optimization

0 引言

近年来，在全球遥感监测中出现许多由多颗遥感卫星构成的星座，其能够实现覆盖全球的高分辨率对地观测。（剩余24190字）