Session
Weekend Session 6: Advanced Concepts - Research & Academia III
Location
Utah State University, Logan, UT
Abstract
On September 3rd 2020, one of the first small satellites equipped with Edge AI hardware was launched. The inclusion of a UB0100 board on PhiSat-1 enabled the use of deep neural networks to provide real-time image analysis on-board an Earth Observation satellite. The primary benefit of this was a 90% reduction in downlink data as the system only transmitted non-cloudy, and thus usable, data. PhiSat-1 and missions like it have started the revolution of satellite-based machine learning, leading ESA and other space agencies to further explore the in-situ deployment of machine-learning models. Other applications that can benefit from on-board space-based machine learning capabilities range from anomaly detection and prognostics to feature recognition and object detection.
This paper focuses on the application of anomaly detection models on space-ready Edge AI hardware to detect and classify anomalous behaviour in telemetry data. The ability to accurately detect anomalies onboard satellite systems has the potential to both increase system lifetimes and reduce satellite operator workloads. The limitations of Edge AI boards and the space environment put restrictions on the models that can be used. Limited power and potential single event upsets constrain the complexity of the models that can be deployed. Therefore, this paper is targeted at models that will run efficiently within these constraints.
We describe an experiment that evaluates the suitability of different anomaly detection approaches (multi-layer-perceptrons, auto-encoders, etc.) for space applications. These approaches are compared both in terms of their performance in the anomaly detection tasks and how well they run on “space ready” low-power hardware. We focus on the Intel Myriad chipset, the basis of the UB0100, which hosted the machine learning image analysis model on PhiSat-1. Our evaluations use both the MIMII machine audio dataset, a well-regarded anomaly detection dataset that is a good proxy for telemetry data, and a dataset generated using anonymized NASA mission telemetry data. The findings show how well basic models work when presented with anomalous satellite telemetry.
Developing Machine Learning Models for Space Based Edge AI Platforms
Utah State University, Logan, UT
On September 3rd 2020, one of the first small satellites equipped with Edge AI hardware was launched. The inclusion of a UB0100 board on PhiSat-1 enabled the use of deep neural networks to provide real-time image analysis on-board an Earth Observation satellite. The primary benefit of this was a 90% reduction in downlink data as the system only transmitted non-cloudy, and thus usable, data. PhiSat-1 and missions like it have started the revolution of satellite-based machine learning, leading ESA and other space agencies to further explore the in-situ deployment of machine-learning models. Other applications that can benefit from on-board space-based machine learning capabilities range from anomaly detection and prognostics to feature recognition and object detection.
This paper focuses on the application of anomaly detection models on space-ready Edge AI hardware to detect and classify anomalous behaviour in telemetry data. The ability to accurately detect anomalies onboard satellite systems has the potential to both increase system lifetimes and reduce satellite operator workloads. The limitations of Edge AI boards and the space environment put restrictions on the models that can be used. Limited power and potential single event upsets constrain the complexity of the models that can be deployed. Therefore, this paper is targeted at models that will run efficiently within these constraints.
We describe an experiment that evaluates the suitability of different anomaly detection approaches (multi-layer-perceptrons, auto-encoders, etc.) for space applications. These approaches are compared both in terms of their performance in the anomaly detection tasks and how well they run on “space ready” low-power hardware. We focus on the Intel Myriad chipset, the basis of the UB0100, which hosted the machine learning image analysis model on PhiSat-1. Our evaluations use both the MIMII machine audio dataset, a well-regarded anomaly detection dataset that is a good proxy for telemetry data, and a dataset generated using anonymized NASA mission telemetry data. The findings show how well basic models work when presented with anomalous satellite telemetry.
