Source-linked AI summary
Structure-Aware Unsupervised Anomaly Detection for Spacecraft Telemetry with Adaptive EVT Thresholding
Óscar Alcarria, Rafael Sánchez, Javier Sempere, Pablo Torrijos, Juan C. Alfaro, Juan M. Auñón, José A. Gámez, José M. Puerta
TL;DR
Spacecraft telemetry anomaly detection often lacks labeled fault histories and extended warm-up data. The paper proposes an unsupervised framework combining mission-aware model selection, incremental monthly retraining, and adaptive EVT thresholding, achieving F0.5 = 0.700 on Mission 1 and F0.5 = 0.698 on Mission 2 under strict chronological evaluation.
Problem
Spacecraft telemetry anomaly detection rarely has complete labeled fault histories, dedicated GPU hardware, or years of telemetry before monitoring begins.
Method
The framework selects models from mission-level statistical properties, retrains incrementally each month, and applies adaptive EVT-based thresholding without labels or extensive historical data.
Results
F0.5 = 0.700 on Mission 1 and F0.5 = 0.698 on Mission 2 under the reported evaluation.
Takeaways & Limitations
The framework matches or surpasses industrial supervised baselines on ESA-AD without labels, GPU hardware, or extensive historical data.
Takeaways & Limitations
Systematic search for model architectures, hyperparameters, and thresholds is identified as future work to reduce manual-tuning bias.
Abstract
from arXiv · showhide
Operational anomaly detection in spacecraft telemetry typically requires labeled historical anomalies or extended warm-up periods. These requirements are rarely met in practice. We propose an unsupervised, deployment-ready framework that produces predictions from the second month of operation without any labels, prior fault knowledge, or mission-specific tuning. The approach combines incremental monthly retraining, statistical model selection, and adaptive Extreme Value Theory (EVT) thresholding for false alarm control. On the ESA Anomalies Dataset (ESA-AD), it achieves $F_{0.5}=0.700$ on Mission~1 and $F_{0.5}=0.698$ on Mission~2 under strict chronological evaluation.
1 Introduction
Spacecraft telemetry anomaly detection must operate on multivariate, non-stationary data despite scarce labeled fault histories and limited operational resources. The proposed structure-aware framework adapts its detection strategy to mission statistics, retrains monthly, and uses adaptive EVT thresholding without labels or extensive historical data.
- Spacecraft telemetry anomaly detection involves multivariate, non-stationary time series under strict operational constraints.
- Mission environments rarely provide complete labeled fault histories, dedicated GPU hardware, or years of telemetry before monitoring begins.
- The framework selects the detection paradigm from mission-level statistical properties, performs monthly retraining, and applies adaptive EVT-based thresholding.
- Without labels, GPU hardware, or extensive historical data, the framework matches or surpasses industrial supervised baselines on ESA-AD.
2 Methodology
The framework targets anomalous events in anonymized multivariate satellite telemetry under realistic deployment constraints. Its workflow combines preprocessing, monthly incremental learning, structure-aware model selection, adaptive EVT thresholding, and anomaly grouping.
- Anomalies are modeled as contiguous off-nominal temporal segments rather than isolated outliers.
- The framework operates without labels or GPU hardware under chronological incremental evaluation to preserve temporal causality.
- The workflow proceeds from preprocessing and monthly incremental learning to structure-aware model selection, adaptive EVT thresholding, and post-detection anomaly grouping.
2.2 Data Preprocessing
Telemetry channels are resampled independently onto mission-specific uniform grids, then aligned into one sparse multivariate series. A channel-aware preservation step retains event-associated values during discretization.
- Asynchronous and irregularly sampled channels are combined over original timestamps to form a sparse multivariate time series.
- Channels are independently projected onto uniform grids with 30 s sampling for M1 and 18 s for M2, using zero-order hold interpolation.
- Event-overlapping intervals copy the last original telemetry value to the following grid timestamp to prevent events disappearing during discretization.
- Resampled channels are aligned on a common temporal grid and concatenated into a single multivariate series.
2.3 Incremental Learning Setup
The incremental setup trains on the previous month and evaluates on the subsequent month. This monthly sliding-window design supports early deployment, progressive adaptation, and operation without large historical datasets.
- The model is trained exclusively on the previous month’s data and evaluated on the subsequent month.
- The setup enables deployment from the first available period without requiring large historical datasets.
- Monthly updates allow the detector to adapt progressively to changing telemetry behavior and align with production monitoring cycles.
2.4 Mission Characterization
Mission 1 and Mission 2 exhibit substantially different statistical structures, with M2 showing stronger clustering and M1 weaker intrinsic segmentation. These differences motivate mission-specific model selection rather than a single detector.
- Distributional differences: Energy Distance between M1 and M2 is 4.33, with a permutation-test p < 0.005 indicating distinct multivariate distributions.The test used 200 permutations while preserving sample sizes.
- Clustering structure: M2 has consistently higher Silhouette coefficients than M1, indicating more compact and better-separated clusters.The comparison evaluates K-means clustering on standardized features across multiple k values.
- Low-dimensional structure: PCA projections reinforce the distinction between missions, with the first two components explaining 33.35% and 19.14% of total variance.PCA was fitted on M1 and then applied to both missions.
- Mission-specific structure: M2 exhibits more separated dense regions and isolated low-density areas, consistent with stronger local clustering structure.M1 shows weaker segmentation by comparison.
- Model-selection implication: The analysis motivates structure-driven model selection because generic standalone detectors performed poorly in preliminary experiments.The paper hypothesizes subtle deviations in M1 and clearer structural separations in M2.
2.5 Model Selection and Training
The framework selects different unsupervised detectors for the missions based on their statistical structure. M1 uses an LSTM autoencoder for subtle temporal deviations, while M2 uses PCA, K-means, and LOF for structured operational regimes.
- M1 model: M1 uses an unsupervised LSTM autoencoder because its low variability and weak clustering suggest subtle temporal deviations.The autoencoder captures temporal dynamics and nonlinear relationships among variables.
- M1 scoring: The M1 autoencoder reconstructs input sequences and quantifies anomaly scores using mean squared error.The model reconstructs each input sequence as x̂_t before computing reconstruction error.
- Training constraints: Model capacity is constrained for CPU-only monthly retraining using sequences of length L = 64, one layer, 64 hidden units, and a 16-dimensional latent space.Training uses Adam with learning rate 10^-3 for up to 30 epochs and early stopping.
- M2 model: M2 uses PCA, K-means, and LOF to model recurrent operational regimes and local deviations within structured state spaces.The anomaly score weights LOF novelty at 0.7 and centroid distance at 0.3.
- Unsupervised calibration: Hyperparameters are inferred from each training window without supervised tuning, including PCA dimensionality explaining 95% cumulative variance and data-dependent cluster selection.The number of clusters is estimated with the Gap Statistic, while LOF neighborhood size scales with √n and reduced dimensionality.
2.6 EVT-Based Dynamic Thresholding
The thresholding procedure adapts anomaly-score cutoffs to changing upper-tail distributions. It combines peaks-over-threshold EVT, fallback quantiles, smoothing, and bounded threshold updates.
- Adaptive EVT: EVT adapts thresholds to the upper tail of each calibration-score distribution when score distributions vary across temporal windows.The procedure uses a peaks-over-threshold approach.
- Tail modeling: For window k, scores above the empirical quantile uk = Qq(Sk) are fitted with a Generalized Pareto Distribution when enough excesses are available.The excesses are fitted by maximum likelihood using shape ξk and scale σk > 0.
- Fallback rule: If fewer than Nmin excesses are available or the GPD fit is invalid, the empirical quantile Q1−α(Sk) provides the fallback threshold.This fallback avoids relying on an invalid or insufficient tail fit.
- Temporal stabilization: Thresholds are exponentially smoothed and their relative changes are limited to δ between consecutive windows.The smoothed threshold combines the current EVT threshold with the previous threshold using λ.
- Decision rule: A score is anomalous when sk,t > Tk, using fixed a priori settings q = 0.98, α = 0.01, Nmin = 30, and λ = δ = 0.05.Here α is a target upper-tail exceedance probability, not a guaranteed false-positive rate.
2.7 Post-detection Anomaly Grouping
Detected anomalies separated by less than six hours are consolidated into one event. This operational grouping reduces redundant alarms and improves diagnostic clarity.
- Event consolidation: Anomalies separated by less than 6 hours are consolidated into a single event.The threshold follows expert recommendations from satellite operations centers.
- Operational rationale: Event consolidation reduces redundant alarms and improves diagnostic clarity.The grouping reflects operational practice rather than isolated point-level reporting.
3 Results
The framework achieves strong event-wise performance under chronological evaluation, with mission-dependent precision–recall tradeoffs and temporal behavior. Results are compared with unsupervised and supervised baselines under differing evaluation configurations.
- Global performance: F0.5 = 0.700 on M1 and F0.5 = 0.698 on M2 summarize the framework’s global event-wise performance.M1 shows high precision with moderate recall, while M2 has a more balanced tradeoff.
- Temporal evolution: M1’s cumulative corrected F0.5 exceeds 0.9 during the initial dense period before stabilizing between 0.7 and 0.8.Anomalous and rare events concentrate heavily in 2000 and become sparser afterward.
- Temporal evolution: M2’s cumulative F0.5 drops briefly during the first months, then recovers quickly and stabilizes around 0.7 with narrower oscillations.M2 event classes remain more constant throughout the mission.
- Baseline comparison: Compared with an industrial unsupervised baseline, the framework improves M1 from F0.5 = 0.424 while maintaining competitive M2 performance.The baseline used static train/test configurations, whereas the framework used incremental monthly retraining without mission-specific hyperparameter tailoring.
- Baseline comparison: Against static 50/50 benchmarks, the framework narrows the gap with supervised methods without labels or channel selection.The cited comparison is not directly comparable because the evaluation configurations differ.
- Computational cost: M1 evaluation required 46.64 hours in total, whereas M2 evaluation required 2.5 hours under CPU-only execution and 16 GB of RAM.Peak memory averaged 7.72 GB for M1 and 0.54 GB for M2, with no inter-window memory accumulation.
4 Discussion
The discussion clarifies how the unsupervised framework handles operational constraints and separates exploratory analysis from deployment decisions. It also notes a possible extension when labeled historical data become available.
- Operational assumptions: Complete-mission exploratory analysis is used only to describe statistical differences and justify model-family choices, not for tuning, calibration, or label-dependent decisions.Operational use may begin with a short warm-up phase before restricting statistical assessment to observed data.
- Operational assumptions: Hyperparameters are selected using standard practices rather than supervised optimization, consistent with the fully unsupervised deployment scenario.This preserves the intended setting in which labeled historical data are unavailable.
- Possible extension: With labeled historical data, AutoML could jointly optimize architectures, hyperparameters, and thresholds using event-wise F0.5.The proposed extension is intended to support adaptive, reproducible model selection and reduce manual-tuning bias.