Virtualization Technology News and Information
Incremental Deep Learning For Satellite with KubeEdge and MindSpore

By Baoyue from KubeEdge and Xiaoman Hu, Zhipeng Huang from MindSpore

This article consists of five parts:

  1. Status Quo and Challenges of LEO Satellites
  2. Tiansuan Constellation Program
  3. KubeEdge Application in the Program
  4. MindSpore Integration With KubeEdge For Cloud Native AI
  5. Future Works

Status Quo and Challenges of LEO Satellites

LEO Satellites has been the recent focus of space technology development, more and more smallsat and cubesat have been launched into the low earth orbit to form constellations, with SpaceX's starlink program as an example.

Despite its promising future, key challenges lies ahead for LEO Satellites:

  • Constellation Management

Automatic satellite tracking and failure detection for satellite constellation to avoid collision and maintain the correct geo-position.

  • Communication Management

On-board automation to minimize orbit-earth communication is needed due to the power consumption limit and radio frequency scarcity. Real time data synchronization is also becoming more prevalent that requires the Satellite always-on 24/7 for data transmission. Real-time in-orbit processing is also required to clear redundant and useless data, shorten the response time, and reduce the pressure or workload of network link transmission

Services such as remote sensing, emergency response, and disaster forecasting are eager for in-orbit computing and processing to improve the response speed and prediction accuracy

  • Space Traffic Management

Satellite health checking, debris avoidance and collision monitoring becomes even more important between different constellations

  • New Application Scenarios

Use LEO Satellites to address the problem of climate change, space mining, deep space learning and so forth.

Tiansuan Constellation Program

To meet these requirements, the Shenzhen Institute of BUPT established the Inter-Planetary Network and Intelligent Computing Lab in June 2020. The lab conducts satellite-related researches, including inter-planetary networks, satellite networks, and distributed AI computing. In March 2021, the lab completed the encapsulation of the satellite-ground collaboration platform.

In July, the lab encapsulated the in-orbit intelligent computing and service platform. These two platforms, along with KubeEdge, were integrated into the satellite launched at Jiuquan Satellite Launch Center on December 7, 2021.

In August 2021, the lab deployed a lightweight 5G core network on a satellite for the first time, realizing interconnection and communication with terrestrial 5G private networks at the signaling layer, as well as traffic splitting for video calls and edge computing.

In September 2021, the lab started to develop the satellite, BUPT 1..

In October 2021, the lab initiated the Tiansuan Constellation program with Spacety Co., Ltd.

Combine the power of cloud native and AI

The objective is to build a constellation as the foundation of intelligent and comprehensive digital infrastructure to support the development of technologies and implement an open-source in-orbit platform of aerospace computing.

By combining KubeEdge and specifically its Sedna module, the Tiansuan-1 Satellite from the Tiansuan Constellation program have the capability of orbit-earth coordination. It enables new service scenarios like orbit-earth coordinated image inference, incremental deep learning and federated learning.

By further combining MindSpore, a new open source deep learning framework on par with TensorFlow and PyTorch, and its unique feature that serves all scenarios including mobile, edge and cloud with just one framework, Tiansuan Constellation is able to have the first in its kind a distributed deep space learning capability.

Adoption of KubeEdge in the Program

KubeEdge is an open source project launched on Nov 2018, which extends native containerized application orchestration capabilities to hosts at Edge. It's built upon Kubernetes and provides fundamental infrastructure support for network, application deployment and metadata synchronization between cloud and edge.

It aims to resolve three major challenges for edge computing: network reliability and bandwidth limit between cloud and edge, resource constraint at edge, highly distributed and heterogeneous device managements.

Its development history is as follows:


From 2018 to 2019, the development of KubeEdge focused on its capabilities. In 2020, the focus was to collaborate with peripheral ecosystems. In 2021, some in-depth attempts were made in a wide range of fields, including AI, robotics, wireless, and MEC. KubeEdge now is an incubating project of Cloud Native Computing Foundation (CNCF), with more than 70 organizations and 950 contributors (including 250 code contributors) around the world.

KubeEdge-Based Cloud-Native Satellite Computing Platform

1. Sedna, an edge collaboration AI sub-project of KubeEdge, is used to build multi-model collaborative inference between the ground and satellite, and incremental model training on the ground based on KubeEdge. In the way of using small models on the satellite and large models on the ground, better support for AI inference can be realized with very few resources on the satellite.

2. The device mapper of KubeEdge is used to model and manage sensors of the satellite in a unified manner, allowing management personnel on the ground to obtain the working status of on-board devices in real time. All of these communicate with each other through a highly reliable cloud-edge channel established by KubeEdge. A Kubernetes data model is also used to realize unified lifecycle management of applications on the satellite through KubeEdge.


Adoption of MindSpore in the Program


MindSpore, a newly open sourced deep learning framework, was launched on March 2020. The main mission is to close the gap between AI research and industry by providing a unified framework for mobile, edge and cloud. Moreover with features such as high-order differentiation optimization, automatic parallelization, and graph operator fusion, MindSpore could achieve very high performance. AI For Science is another focus of MindSpore which provides avant-garde deep learning optimization with traditional complex numerical scientific computing like the computation of Navier-Stoke, Maxwell and Schrödinger equations.

Since its inception, MindSpore established an open and global community of developers. In merely two years, MindSpore has achieved more than 1.2 million downloads with more than 20 MindSpore Study Groups created around the world. The community adopt open governance with a 14 member Technical Steering Committee and 20 SIGs/WGs. The community has also pioneered in equality and diversity with efforts like MSG-Women In Tech. With various community partners, non-for-profit collaboration like pre-trained ultra-red camera model for natural protection further helps AI For Good. The community team also developed a high level API tookit for MindSpore called TinyMS aiming for non-AI centric service's fast adoption of deep learning framework capability.

KubeEdge-MindSpore Integration for Cloud Native AI

  1. MindSpore yolov3-tiny model is predeployed on the satellite and served with KubeEdge-Sedna via TinyMS (with a size of 30MB, which fits in with the satellite memory).
  2. Upon discovering “hard samples”, KubeEdge-Sedna LC (Local Controller) compresses and sends these datasets to the ground (with 100Mbps downlink). Then KubeEdge-Sedna GM (Global Manager)performs incremental learning tasks using MindSpore to train new AI models and improve AI algorithm precision.
  3. After refreshing the model that is finetuned and retrained by MindSpore, KubeEdge-Sedna pushes the compressed partial model (with a size of 3 MB, which fits in with the 1Mbps uplink) to the satellite.
  4. After receiving the compressed partial model, MindSpore decompresses and updates the model, and then KubeEdge-Sedna redeploys the new model.

    Future plans: deep space learning and cloud native space

    KubeEdge and MindSpore have many other amazing capabilities, and there are more exciting new scenarios in the future:

    • Deep Space Learning for Earth-Orbit/Mars-Orbit/Venus-Orbit missions that facilitate automatic constellation forming or smallsat/cubesat fleet management
    • Cloud Native Policy for multi-tenant ride-sharing smallsat/cubesat resource isolation and security/privacy
    • Cloud Native Space Computing that enables better Orbit-Ground or Orbit-Orbit communication, monitoring and resource management.

    You are welcomed to join the community and participate in experiments and discussions on satellite in-orbit computing and aerospace computing.

    More information for KubeEdge and MindSpore:


    ***To learn more about containerized infrastructure and cloud native technologies, consider joining us at KubeCon + CloudNativeCon Europe 2022, May 16-20. 



    Yue Bao serves as a software engineer of Huawei Cloud. She is now working 100% on open source, focusing on lightweight edge and edge api-server for KubeEdge. Before that, Yue worked on Huawei Cloud Intelligent EdgeFabric Service and participated in multiple edge engineering projects. Yue graduated from Chu Kochen Honors College of Zhejiang University, majoring in information science and electronic engineering.


    Xiaoman Hu is operation director of MindSpore in Huawei. Member of the Chinese Institute of Electronics experts, member of the Outreach Committee of the LF&AI Foundation, leader of the TinyMS open source project. Founder of MSG•Women in Tech. Served as a senior algorithm engineer of Autohome, she built a distributed deep learning framework from scratch, and led multiple image recognition and object detection projects. In addition, she served as a deep learning evangelist for Baidu, responsible for the PaddlePaddle developer ecosystem, and built a content ecosystem from scratch.


    Zhipeng Huang currently serves as open source director for Huawei Ascend product. Zhipeng is now the TAC member of LFAI, TAC and Outreach member of the Confidential Computing Consortium, co-lead of the Kubernetes Policy WG, project lead of CNCF Security SIG, founder of the OpenStack Cyborg project, and co-chair of OpenStack Public Cloud WG. Zhipeng is also leading a team in Huawei that works on ONNX, Kubeflow, Akraino, and other open source communities."

    Published Friday, May 06, 2022 7:31 AM by David Marshall
    There are no comments for this post.
    To post a comment, you must be a registered user. Registration is free and easy! Sign up now!
    <May 2022>