Session
Science/Mission Payloads
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
Onboard machine learning is an increasingly important topic in spaceflight systems, as machine learning and AI models are increasingly used for data processing, anomaly detection, attitude control systems, and other predictive applications. However, successfully deploying a machine learning model in a spaceflight computing environment is fraught with difficulty, as spaceflight computing environments are low-resource, but typical machine learning software solutions are implemented in languages like Python, which require an interpreter and other heavyweight dependencies. It is therefore virtually mandatory for any on-spacecraft machine learning to be written with lightweight toolkits that minimize dependencies and resource usage.
We demonstrate a solution to this problem: via the use of open source C++-based scientific computing software, including the mlpack machine learning library and Armadillo linear algebra library, we provide the first implementation of a working on-spacecraft machine learning anomaly detection system inside the F Prime flight software framework that successfully detects non-trivial anomalous events by using a kernel density estimate on telemetry data. Our prototype is validated on real hardware, where we physically created several anomalous situations that were all successfully detected. The code is available publicly on Github and can be used as a reference project for deploying any complex machine learning pipeline to a spacecraft via the F Prime framework.
Document Type
Event
Lightweight Open Source On-Spacecraft Machine Learning With mlpack and F Prime
Salt Palace Convention Center, Salt Lake City, UT
Onboard machine learning is an increasingly important topic in spaceflight systems, as machine learning and AI models are increasingly used for data processing, anomaly detection, attitude control systems, and other predictive applications. However, successfully deploying a machine learning model in a spaceflight computing environment is fraught with difficulty, as spaceflight computing environments are low-resource, but typical machine learning software solutions are implemented in languages like Python, which require an interpreter and other heavyweight dependencies. It is therefore virtually mandatory for any on-spacecraft machine learning to be written with lightweight toolkits that minimize dependencies and resource usage.
We demonstrate a solution to this problem: via the use of open source C++-based scientific computing software, including the mlpack machine learning library and Armadillo linear algebra library, we provide the first implementation of a working on-spacecraft machine learning anomaly detection system inside the F Prime flight software framework that successfully detects non-trivial anomalous events by using a kernel density estimate on telemetry data. Our prototype is validated on real hardware, where we physically created several anomalous situations that were all successfully detected. The code is available publicly on Github and can be used as a reference project for deploying any complex machine learning pipeline to a spacecraft via the F Prime framework.
