Session
Weekday Poster Session 4
Location
Utah State University, Logan, UT
Abstract
Small satellites in Low Earth Orbit (LEO) constellations are shaping the future of communications and Earth observation. Despite having inherent advantages such as lower latency and faster deployments due to quicker and cheaper access to space, monitoring the large number of LEO satellites imposes a significant burden when considering the long-established methods of human-in-the-loop anomaly detection and recovery. Timely responses to anomalies reduce operational outages and help maximize the availability of the network. Traditional rule-based detection and supervised learning have inherent limitations in monitoring the large numbers of space assets projected to be launched over the next decades because more human intervention will be required to ensure anomaly detection. Although studies demonstrate the potential of unsupervised learning in detecting system anomalies using time-series telemetry data from space missions, such studies are limited to small telemetry datasets (i.e. less than 100 mnemonics). Because it is common for satellites to have many hundreds or thousands of mnemonics, adapting such Machine Learning (ML) models for time-series anomaly detection at a larger scale remains challenging. We examined the TelemAnom model developed by NASA using years of historical satellite telemetry data and proposed improvements to the TelemAnom model for scalable time-series anomaly detection on operational platforms. We evaluated the predictability of our adapted model on empirical telemetry data and compared model-identified anomalies with known anomalies identified by subject matter experts. Our finding suggests that factors such as satellite design, availability of data, and Exploratory Data Analysis (EDA) processes are important considerations when aligning unsupervised models with traditional time-series anomaly detection methods.
Prompt Anomaly Detection for Small Satellites in Low-Earth Orbit Constellations: A Machine Learning Approach
Utah State University, Logan, UT
Small satellites in Low Earth Orbit (LEO) constellations are shaping the future of communications and Earth observation. Despite having inherent advantages such as lower latency and faster deployments due to quicker and cheaper access to space, monitoring the large number of LEO satellites imposes a significant burden when considering the long-established methods of human-in-the-loop anomaly detection and recovery. Timely responses to anomalies reduce operational outages and help maximize the availability of the network. Traditional rule-based detection and supervised learning have inherent limitations in monitoring the large numbers of space assets projected to be launched over the next decades because more human intervention will be required to ensure anomaly detection. Although studies demonstrate the potential of unsupervised learning in detecting system anomalies using time-series telemetry data from space missions, such studies are limited to small telemetry datasets (i.e. less than 100 mnemonics). Because it is common for satellites to have many hundreds or thousands of mnemonics, adapting such Machine Learning (ML) models for time-series anomaly detection at a larger scale remains challenging. We examined the TelemAnom model developed by NASA using years of historical satellite telemetry data and proposed improvements to the TelemAnom model for scalable time-series anomaly detection on operational platforms. We evaluated the predictability of our adapted model on empirical telemetry data and compared model-identified anomalies with known anomalies identified by subject matter experts. Our finding suggests that factors such as satellite design, availability of data, and Exploratory Data Analysis (EDA) processes are important considerations when aligning unsupervised models with traditional time-series anomaly detection methods.
