Related Work

Efficient network delivery for big in-orbit data. Emerging satellites with evolved remote sensing capabilities are widely used in many applications such as the earth surveillance and disaster monitoring. A number of recent efforts have studied the approaches for accelerating space data delivery and optimizing the task completion time [12], [40], [41]. L2D2 [12] is a space data download scheme which uses commodity hardware to offer low latency and robust download. OrbitCast [40] is a hybrid space data delivery architecture that collaboratively leverages LEO satellites and geo-distributed ground stations to fast forward space data. However, with the increasing resolution of emerging on-board sensors, the amount of space data has also increased exponentially in recent years. Downloading all space data to the ground requires massive amounts of bandwidth and storage in satellite, which induces large downloading latency.

高效的在轨大数据网络传输。 随着遥感能力不断提升的新型卫星，已广泛应用于地球观测、灾害监测等诸多场景。近年来，一些研究工作致力于加速空间数据传输、优化任务完成时间 [12], [40], [41]。例如，L2D2 [12] 是一种空间数据下载方案，利用商用硬件实现低时延且鲁棒的下载性能；OrbitCast [40] 则是一种混合式空间数据传输架构，通过协同利用低轨（LEO）卫星与地理分布的地面站，加速空间数据下传。然而，随着新型星载传感器分辨率的不断提高，近年来空间数据量呈指数级增长。将所有空间数据下载至地面不仅需要大量卫星带宽与存储资源，还会带来显著的下载延迟。

SEC for in-orbit data processing. Other efforts investigated the feasibility of leveraging edge-like computation capabilities on emerging satellites to directly process data in orbit [7], [8]. This technique, known as space/orbit edge computing, can identify and discard unnecessary information among the big space data. Network efficiency is improved since only high-value data will be downloaded to the ground. However, these works ignore the energy challenge caused by the additional workload of in-orbit processing. More recently, MHSPO [9] leverages Lyapunov optimization to optimize energy consumption for SEC networks by offloading tasks to peer satellites. The fundamental difference between MHSPO and PHOENIX is that MHSPO ignores the time-varying sunlight states of LEO satellites, which weakens the effectiveness of battery energy optimization for SEC networks.

面向在轨数据处理的空间边缘计算（SEC）。 另一部分研究工作探讨了在新型卫星上引入类边缘计算能力以直接在轨处理数据的可行性 [7], [8]。这种技术被称为空间/在轨边缘计算（Space/Orbit Edge Computing），可在大规模空间数据中识别并丢弃无用信息，从而提高网络传输效率，仅将高价值数据下传至地面。然而，这些工作忽视了在轨处理额外工作负载所带来的能耗挑战。近期，MHSPO [9] 通过利用 Lyapunov 优化，将任务卸载至对等卫星，从而优化 SEC 网络能耗。MHSPO 与 PHOENIX 的根本区别在于，前者忽视了低轨卫星受光状态的时变特性，从而削弱了电池能耗优化在 SEC 网络中的效果。

Task scheduling in mobile edge computing. Energy-efficient task scheduling [10], [11], [42], [43] has been well studied in terrestrial networks, which exploit dynamic CPU frequency scaling, network interface selection and transmission power allocation techniques to adaptively make offloading decisions and control energy consumption. However, the large amount of tasks makes them difficult to scale the satellite computation and communication capability. Recent learning-based works [44], [45] have applied deep reinforcement learning in mobile devices to select offloading actions. However, the action space explodes among large-scale satellite agents and selecting actions via deep neural network consumes extra energy.

移动边缘计算中的任务调度。 节能型任务调度 [10], [11], [42], [43] 已在地面网络中得到广泛研究，这些方法利用动态 CPU 频率调节、网络接口选择、传输功率分配等技术，来自适应地做出卸载决策并控制能耗。然而，面对海量任务，难以在卫星网络中同时扩展计算与通信能力。近年来，一些基于学习的方法 [44], [45] 尝试将深度强化学习应用于移动设备，以选择卸载动作。但在大规模卫星网络中，动作空间呈爆炸式增长，同时通过深度神经网络选择动作也会带来额外能耗。