Session
Weekend Session 1: Advanced Concepts - Research & Academia I
Location
Utah State University, Logan, UT
Abstract
Autonomy can increase reaction speed, flexibility, and accuracy of satellite operations, especially in uncertain environments caused by delayed communication and/or adversarial conditions. An increased focus on small satellites makes the development of satellite autonomy even more salient, given fewer operators per satellite.
Anomaly detection automates satellite health monitoring, ensuring it functions as designed. This is typically achieved using various forms of recurrent neural networks (RNN). While many of these model-based works show promise, a majority use simulated data or assume lossless communication. In contrast, raw satellite telemetry often has dropped packets, sampling frequency mismatches, noise from electrical systems and radiation, and a lack of clear labels for training.
This work demonstrates how data-centric artificial intelligence (AI) can be utilized in satellite autonomy, using telemetry from the Very Low Frequency Propagation Mapper (VPM) small satellite flown by the Air Force Research Lab Space Vehicle Directorate in 2020. We introduce simple, but effective, tools for extracting fault labels from system parameters, resampling outliers to a common, uniform timeline, and evaluating outlier fault predictability. Results find that detected outliers were able to predict faults 1-10 minutes before they occurred with high accuracy.
Methods for Data-centric Small Satellite Anomaly Detection and Fault Prediction
Utah State University, Logan, UT
Autonomy can increase reaction speed, flexibility, and accuracy of satellite operations, especially in uncertain environments caused by delayed communication and/or adversarial conditions. An increased focus on small satellites makes the development of satellite autonomy even more salient, given fewer operators per satellite.
Anomaly detection automates satellite health monitoring, ensuring it functions as designed. This is typically achieved using various forms of recurrent neural networks (RNN). While many of these model-based works show promise, a majority use simulated data or assume lossless communication. In contrast, raw satellite telemetry often has dropped packets, sampling frequency mismatches, noise from electrical systems and radiation, and a lack of clear labels for training.
This work demonstrates how data-centric artificial intelligence (AI) can be utilized in satellite autonomy, using telemetry from the Very Low Frequency Propagation Mapper (VPM) small satellite flown by the Air Force Research Lab Space Vehicle Directorate in 2020. We introduce simple, but effective, tools for extracting fault labels from system parameters, resampling outliers to a common, uniform timeline, and evaluating outlier fault predictability. Results find that detected outliers were able to predict faults 1-10 minutes before they occurred with high accuracy.
