北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2023, Vol. 46 ›› Issue (2): 57-63.

• 算力网络与分布式云 • 上一篇    下一篇

低轨卫星网络星载边缘DNN推理策略研究

谢人超1,1,1,杨煜天2,唐琴琴1,陈清霞3,向雪霜3   

  1. 1. 北京邮电大学
    2. 北京邮电大学网络与交换技术国家重点实验室
    3. 中国空间技术研究院
  • 收稿日期:2022-09-28 修回日期:2022-09-23 出版日期:2023-04-28 发布日期:2023-05-14
  • 通讯作者: 唐琴琴 E-mail:qqtang@bupt.edu.cn

Reaserch of On-Board Edge DNN Inference Strategies For LEO Satellite Networks

  • Received:2022-09-28 Revised:2022-09-23 Online:2023-04-28 Published:2023-05-14

摘要: 随着低轨卫星在轨计算能力的提升与人工智能服务例如目标探测、卫星侦察等服务需求量激增,深度神经网络(DNN)以其独特的模型结构与高效的学习成为实现智能服务的首选。为解决由于卫星总处于高速移动状态、体积小且异构化而带来的资源强受限、通信较困难等问题,通过低轨卫星实现边缘计算并通过DNN进行分布式任务推理成为必然趋势。首先利用了有向无环图(DAG)探索了DNN模型结构,并研究了低轨卫星网络中的分布式DNN推理问题,然后设计了基于激励函数和处理时延的量子进化算法(QEA),实现了采样率设置和任务卸载的最优化决策。最后,通过仿真测试验证了基于激励函数与处理时延的量子进化算法整体性能优于传统方法。

关键词: DNN分布式推理, 低轨卫星, 任务卸载

Abstract: With the improvement of on-orbit computing capability of low-orbit satellites and the surge in demand for artificial intelligence services such as target detection and satellite reconnaissance, deep neural network (DNN) has become the first choice for realizing intelligent services with its unique model structure and efficient learning. In order to solve the problems of limited resources and difficult communication caused by the satellite always moving at high speed, small size and isomerization, it has become an inevitable trend to realize edge calculation by low-orbit satellite and distributed task inference by DNN. Firstly, directed acyclic graph (DAG) is used to explore the structure of DNN model, and the distributed DNN inference problem in low orbit satellite network is studied. Then, a quantum evolution algorithm (QEA) based on excitation function and processing delay is designed to realize the optimal decision of sampling rate setting and task offloading. Finally, the simulation results show that the performance of quantum evolution algorithm based on excitation function and processing delay is better than that of traditional methods.

Key words: Distributed DNN Inference, LEO Satellite, Task Offloading

中图分类号: