Orbital Edge Computing

OEC is a nanosatellite system design consisting of a set of organizational principles that relies on near-sensor processing in order to avoid the limitations of bent-pipe architectures. We first provide an overview of an individual OEC nanosatellite, i.e. a computational nanosatellite. We then describe a computational nanosatellite pipeline (CNP), which organizes a constellation into a parallel pipeline to hide processing latency by leveraging formation flying techniques [5, 62].

轨道边缘计算（OEC）是一种纳米卫星系统设计，它包含一套依赖近传感器处理的组织原则，旨在避免“弯管”架构的局限性。我们首先概述单个OEC纳米卫星，即计算型纳米卫星。然后，我们描述计算型纳米卫星流水线（Computational Nanosatellite Pipeline, CNP），它利用编队飞行技术[5, 62]将星座组织成一个并行流水线，以隐藏处理延迟。

Computational Nanosatellites

A computational nanosatellite is a nanosatellite with several key changes to its computing hardware and power system.

计算型纳米卫星是在其计算硬件和电源系统上进行了几项关键改变的纳米卫星。

Computing Hardware. A computational nanosatellite supplements existing sensing, communication, and control hardware with onboard computing. In this work, we characterize onboard computing with a Jetson TX2 industrial module. The Jetson TX2 includes a high-capability, low-power, efficient mobile GPU; the industrial variant is designed for extreme temperature environments. Its small volume allows for integration among all other necessary components within the 1U volume available to a 3U cubesat containing a 2U camera system. A 7.5 W power mode closely matches the 7.1 W input power provided by surface-mounted solar panels. The OEC computing model supports computing systems other than the Jetson by varying input performance and energy parameters.

Energy. A computational nanosatellite harvests and stores energy. In this work, the energy harvester is a low-risk, chassis-mounted solar cell array that avoids the mechanical complexity of deployable panels. A chassis-mounted array limits total solar cell area and, as a result, available power peaks at about 7.1 W. A high-density supercapacitor bank stores energy. Supercapacitors hold less total energy than batteries of the same volume, but offer several advantages. Supercapacitors charge quickly and provide immediate, high current; batteries charge slowly and are current-limited. Supercapacitors operate across the wide range of temperatures in space, while batteries fail in excessive heat or cold. OEC systems operate like intermittent systems [16, 17, 37–39, 61, 65, 93, 100], harvesting energy while sleeping to charge capacitors. When energy is sufficient, it performs its sensing, computing, or communication task.

Operating Model. A computational nanosatellite operates by capturing an image and processing it locally instead of transmitting it to Earth through a bent pipe. The application determines the processing method. Examples include CNN-based image classification, object detection, and segmentation, or any other computation; Section 7 evaluates OEC systems with onboard machine inference. A typical OEC processing task identifies features of interest, separating them from raw sensor data. An OEC system discards uninteresting data and sends processed features of interest to Earth, using intelligent early discard as described by prior work [19].

计算硬件。 计算型纳米卫星在现有的传感、通信和控制硬件之外，增加了星上计算能力。在本研究中，我们以一个Jetson TX2工业模块来表征星上计算。Jetson TX2包含一个高性能、低功耗、高效率的移动GPU；其工业版本专为极端温度环境设计。其小巧的体积使其能够在一个3U立方体卫星中（包含一个2U相机系统）可用的1U空间内，与所有其他必要组件集成在一起。其7.5W的功率模式与表面贴装太阳能电池板提供的7.1W输入功率非常匹配。OEC计算模型通过调整输入性能和能量参数，也支持除Jetson之外的其他计算系统。

能源。 计算型纳米卫星负责采集和储存能量。在本研究中，能量采集器是一个低风险的、安装在卫星主体上的太阳能电池阵列，避免了可展开式帆板的机械复杂性。安装在主体上的阵列限制了太阳能电池的总面积，因此可用功率峰值约为7.1W。一个高密度超级电容器组负责储存能量。同等体积下，超级电容器的总储能量低于电池，但具有多项优势。超级电容器充电迅速，并能立即提供高电流；而电池充电缓慢且受电流限制。超级电容器能在太空的宽温度范围内工作，而电池在过热或过冷时会失效。OEC系统的运行方式类似于间歇式系统（intermittent systems）[16, 17, 37–39, 61, 65, 93, 100]，在休眠时采集能量为电容器充电。当能量充足时，它便执行其传感、计算或通信任务。

运行模式。 计算型纳米卫星的运行方式是捕获图像并在本地处理，而不是通过“弯管”架构将其传输到地球 。处理方法由具体的应用决定。例子包括基于CNN的图像分类、目标检测和分割，或任何其他计算；第7节使用星上机器推断来评估OEC系统。 一个典型的OEC处理任务是识别感兴趣的特征，将它们从原始传感器数据中分离出来。OEC系统会丢弃不感兴趣的数据，并将处理过的感兴趣特征发送到地球，这采用了先前工作中描述的智能早期丢弃（intelligent early discard）策略[19]。

Computational Nanosatellite Pipelines

An OEC system is energy and latency constrained. While processing a frame, a nanosatellite can capture but not process additional frames. A nanosatellite cannot capture a frame while sleeping. Effective OEC systems leverage the constellation as a whole to overcome energy and latency constraints.

A constellation of OEC nanosatellites overcomes the energy and latency limitations of individual satellites by organizing them into a computational nanosatellite pipeline (CNP). A CNP leverages existing formation flying techniques [5, 62] to orbit in a fixed configuration, parallelizing data collection and processing across a constellation. CNPs divide image frames into tiles; in some modes, tile processing is distributed among satellites to reduce system-level frame processing latency. Figure 4 illustrates a CNP operating on ground track frames, which are tiled during processing.

一个OEC系统受到能量和延迟的约束。 在处理一帧图像时，一颗纳米卫星可以捕获但无法处理额外的帧。在休眠时，纳米卫星无法捕获帧。有效的OEC系统利用整个星座来克服能量和延迟的约束。

由OEC纳米卫星组成的星座通过将它们组织成一个计算型纳米卫星流水线（CNP）来克服单个卫星的能量和延迟限制。CNP利用现有的编队飞行（formation flying）技术[5, 62]以固定配置在轨运行，从而在整个星座中并行化数据收集和处理。CNP将图像帧分割成“瓦片”（tiles）；在某些模式下，瓦片处理任务被分配给不同的卫星，以降低系统级的帧处理延迟。图4展示了一个在星下点轨迹帧上运行的CNP，这些帧在处理过程中被分割成瓦片。

---

这一部分建议看英文原文:

We identify and evaluate several modes of operation for CNPs:

(1) frame-spaced, tile-parallel;

(2) frame-spaced, frame-parallel;

(3) close-spaced, tile-parallel;

(4) close-spaced, frame-parallel.

Figure 5 illustrates each of these modes. Frame-parallel and tile-parallel describe how image processing tasks are distributed across a CNP.

• Under frame-parallel processing, each nano-satellite processes all tiles in each captured frame.
• Under tile-parallel processing, each nano-satellite processes a subset of tiles in each captured frame.

Frame-spaced and close-spaced describe the physical configuration of a CNP.

• A frame-spaced pipeline places each nano-satellite exactly one GTF apart in distance.
• A close-spaced pipeline places each nano-satellite as close together as is feasible, e.g. meters or tens of meters apart, with a requirement that the end-to-end pipeline distance is less than the length of one GTF.

(1) A frame-spaced, tile-parallel CNP separates devices by one GTF in distance; each device images every GTF (so long as there is sufficient energy) and processes a subset of tiles.

(2) A frame-spaced, frame-parallel CNP also separates devices by one GTF in distance; each device images a distinct subset of GTFs and processes all tiles in the frame.

(3) A close-spaced, tile-parallel CNP places devices close together in distance; each device images every GTF and processes a subset of tiles.

(4) A close-spaced, frame-parallel CNP also places devices close together in distance; each device images a distinct subset of GTFs and processes all tiles in the frame.

(5) An orbit-spaced constellation, in which satellites are evenly distributed across an orbit, is a modified version of a frame-spaced constellation offering improved communication opportunities.

---

Station-keeping, which allows a nanosatellite convoy to maintain consistent distances between adjacent devices (e.g. framespaced, close-spaced), requires formation flying techniques to compensate for atmospheric drag and other astrodynamic effects.

位置保持（Stationkeeping）能让一个纳米卫星编队在相邻设备间保持一致的距离（例如帧间隔、近距离间隔），这需要编队飞行技术来补偿大气阻力和其他天体动力学效应。

The number of devices in a CNP, or pipeline depth, increases the aggregate parallel work performed and the aggregate energy collected. Pipeline depth does not affect total data per revolution, because the number of frames per orbit remains constant. When the aggregate energy harvested by a CNP is less than the aggregate energy required to process all data, adding devices increases coverage (the fraction of GTFs captured per revolution) by increasing total system energy per revolution. When the aggregate energy harvested by a CNP is enough to process all data, a CNP may achieve full coverage. However, such a pipeline may still fall short of full coverage due to processing latency. If there are too few nanosatellites to complete parallel processing of all frames in one revolution, then coverage remains incomplete. Adding devices to the pipeline increases parallelism and decreases latency, eventually yielding a system capable of full coverage.

CNP中的设备数量，即流水线深度（pipeline depth），会增加总体的并行工作量和总体的能量收集量。流水线深度不影响每个轨道周期的总数据量，因为每个轨道的帧数是恒定的。当CNP收集的总能量少于处理所有数据所需的总能量时，增加设备可以通过增加每个轨道周期的系统总能量来提高覆盖率（每个轨道周期捕获的GTF的比例）。当CNP收集的总能量足以处理所有数据时，CNP可能实现完全覆盖。然而，这样的流水线仍可能因处理延迟而无法达到完全覆盖。如果纳米卫星数量太少，无法在一个轨道周期内完成所有帧的并行处理，那么覆盖率仍然是不完整的。向流水线中增加设备可以提高并行度并降低延迟，最终使系统能够实现完全覆盖。

A computational nanosatellite pipeline requires propulsion and positioning. Unlike uncontrolled constellation configurations, formation flying requires each nanosatellite to have a propulsion system to correct for drag. One recent survey describes a variety of nanosatellite propulsion systems [91], including cold gas, liquid, ion thrusters, and hall effect propulsion systems. Additionally, 39 deployed and tested propulsion systems were evaluated in a recent survey of nanosatellite formation flying [5]. The wealth of recent research on propulsion makes CNP formation flying feasible.

To avoid the complications and expense of cross-link satellite communication, each device independently triggers camera captures based on position. The predetermined orbit and formation of a CNP allows capture coordinates to be defined before launch. Contemporary navigation constellation receivers track position with milliwatts of power, and they can be unlocked for high-velocity, high-altitude use in space [9]. We anticipate that cubesat pipelining may motivate further research in nanosatellite guidance, navigation, and control.

计算型纳米卫星流水线需要推进和定位能力。与非受控的星座配置不同，编队飞行要求每颗纳米卫星都具有推进系统以修正阻力。最近的一项综述描述了多种纳米卫星推进系统[91]，包括冷气体、液体、离子推进器和霍尔效应推进系统。此外，最近一项关于纳米卫星编队飞行的综述评估了39个已部署和测试的推进系统[5]。近期在推进系统方面的大量研究使得CNP的编队飞行变得可行。

为避免跨链卫星通信的复杂性和开销，每个设备根据自身位置独立触发相机捕获。CNP预定的轨道和编队允许在发射前就定义好捕获坐标。当代的导航星座接收机能以毫瓦级的功率追踪位置，并且可以被解锁以用于太空中的高速、高海拔环境[9]. 我们预计，立方体卫星的流水线化可能会进一步推动纳米卫星制导、导航与控制领域的研究。