A Comparative Analysis of Forecasting Foundation Models for Spacecraft Multivariate Time Series Anomaly Detection

Alexandre Olive, Sundip R. Desai, Richard H. Foster, Moses Chan, Mary Comer, Edward Delp

Keywords: anomaly detection, spacecraft multivariate time series, forecasting foundation models

TL;DR: We evaluate zero-shot and fine-tuned forecasting foundation models for multivariate spacecraft telemetry anomaly detection.

Abstract: Detecting anomalies in spacecraft telemetry is critical for early warning of potential system failures. Forecasting foundation models have shown success across various domains. However, their effectiveness for time series anomaly detection has not been thoroughly evaluated, especially for spacecraft telemetry data. We explore the use of multiple pre-trained forecasting foundation models for anomaly detection in spacecraft telemetry, including MOIRAI (Masked Encoder-based Universal Time Series Forecasting Transformer), Toto, Chronos and TimesFM. We use these models to forecast multivariate time series and estimate the forecast error between actual and predicted values. We evaluate both zero-shot and fine-tuned approaches, comparing univariate models (Chronos, TimesFM) with multivariate models (MOIRAI, Toto). We evaluate the performance on a Mars Reconnaissance Orbiter (MRO) dataset with multiple telemetry channels and four injected anomaly types. Results show that Toto achieves the best zero-shot performance with an F1-score of 0.752, while fine-tuned MOIRAI achieves the best overall performance with an F1-score of 0.859. Domain-specific fine-tuning and multivariate processing improve anomaly detection performance for spacecraft telemetry.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 47