Insights into the Benchmarking of Adaptive Ensemble Learning for Robust Time-series Anomaly Detection in Satellite Telemetry Data

Abstract

Time-series Anomaly Detection (TSAD) in satellite telemetry is traditionally reliant on Out-Of-Limits (OOL) detection, which suffers from key limitations such as the inability to capture temporal trends or changes in channel frequencies. On the other hand, modern TSAD algorithms may thrive on individual types of anomalies (e.g. contextual) or particular distributions (e.g. multiple similar anomalies) but lack flexibility and fail in out-of-distributions scenarios. To improve the robustness of our TSAD pipeline and enable a root cause analysis, we use an ensemble-based approach that incorporates various families of TSAD models. This method aims at improving the adaptability of our TSAD pipeline across different spacecraft operations.

This work extends an initial benchmark of individual TSAD methods on the ESA Anomaly Detection Benchmark (ESA-ADB) for time-series analysis in satellite telemetry. Our early findings favour the use of out-of-the-box but robust unsupervised methods (e.g. iForest, HBOS) over supervised methods struggling with the highly heterogeneous dataset. Furthermore, our latest analysis shows that adaptive model selection for TSAD can offer a computationally efficient alternative to full ensemble methods by frequently recommending the most suitable TSAD method per subsystem, improving scalability without compromising detection accuracy in large missions.