A Hierarchical Ensemble Pipeline for Anomaly Detection in ESA Satellite Telemetry

Technical Summary: A Hierarchical Ensemble Pipeline for Anomaly Detection in ESA Satellite Telemetry

Problem Statement
The paper addresses the challenge of detecting anomalies in multivariate telemetry data from European Space Agency (ESA) satellites. Satellite operations occur in hostile environments where timely anomaly detection is critical to prevent operational disruptions and economic loss. The specific challenge involves monitoring 58 out of 76 telemetry channels simultaneously. The data presents significant statistical difficulties, including irregular sampling, missing values, channel redundancy, and a low anomaly density (1.80%). Furthermore, the dataset contains "nominal rare events" (e.g., resets, calibrations) that are statistically abnormal but not faults, requiring models to distinguish between true anomalies and rare nominal behaviors. The evaluation metric, the corrected event-wise F0.5 score, emphasizes precision and the ability to detect entire anomaly intervals rather than individual samples.

Methodology
The authors propose the ESA Hierarchical Ensemble Pipeline, a modular framework integrating statistical segmentation, shapelet-based representation, and multi-level ensemble learning. The architecture is designed to prevent information leakage and ensure model diversity through a specific data handling strategy.

1. Two-Level Masking Strategy:
   To ensure independence between layers and prevent data leakage, the authors employ a two-level masking strategy during training. The input time series is split into $N$ segments. For each configuration, the $n$-th segment is masked at the first level, and a sub-segment is masked at the second level.

   • Level 1: The remaining data ($\hat{x}_n \setminus \hat{x}_{n,m}$) is used for base model training and shapelet extraction.
   • Level 2: The masked segment $x_n$ is used exclusively to train the intra-channel stacking meta-model.
   • Cross-Channel: A separate segment ($x_{cca}$) cut from the end of the raw series is reserved for training the cross-channel aggregation model.
     This ensures that no training information is reused across layers, maintaining feature independence.

2. Preprocessing and Feature Extraction:

   • Segmentation: Sliding-window segmentation is applied with fixed duration and stride.
   • Feature Extraction: Two types of features are extracted:
     • Base Statistical Features: Mean, variance, skewness, kurtosis, STFT energy, spectral centroid, slope, and diff_var, selected via XGBoost importance and visual analysis.
     • Shapelet Features: Local patterns (shapelets) are mined from anomalous segments within the masked training data. These shapelets are used to compute similarity scores (maximum and minimum) against input segments, capturing discriminative local patterns indicative of anomalies.
   • Pooling: A rolling min-max pooling is applied to the feature timeline. This reduces redundancy and emphasizes extreme feature responses, which is crucial for the event-wise F0.5 metric where a single high-confidence detection within an interval suffices.

3. Hierarchical Ensemble Architecture:

   • Layer 1 (Base Models): Independent classifiers (XGBoost) are trained on each channel using unique data slices and shapelet dictionaries derived from disjoint data portions.
   • Layer 2 (Intra-Channel Stacking): A meta-classifier (Logistic Regression) aggregates predictions from multiple base models for each channel. It is trained on the masked segment $x_n$ to avoid bias from the base models' training data.
   • Layer 3 (Cross-Channel Aggregation): This layer integrates outputs from semantically coherent groups of channels. A "group reducer" function (power-weighted sum based on validation precision) aggregates channel probabilities. These group-level signals are fed into a final multivariate model (tested: XGBoost, Logistic Regression, LSTM) trained on the $x_{cca}$ segment.

Key Contributions

• Hierarchical Ensemble Pipeline: A novel framework combining statistical and shapelet-based feature extraction with a three-level ensemble structure (base, intra-channel stacking, cross-channel aggregation).
• Two-Level Masking Strategy: A rigorous data splitting technique that prevents information leakage across hierarchy levels and promotes model diversity by forcing models to learn from distinct data views.
• Validation on ESA-ADB: The method was rigorously tested on the ESA Anomaly Detection Benchmark (ESA-ADB) challenge, demonstrating strong generalization on realistic satellite telemetry.

Results
The pipeline was evaluated on the ESA-ADB challenge using the corrected event-wise F0.5 score.

• Performance: The proposed algorithm achieved 1st place on the public leaderboard (score: 0.931) and 3rd place on the private leaderboard (score: 0.853).
• Model Configuration:
  • Aggregator: Logistic Regression proved to be the most robust and effective model for the cross-channel aggregation layer, outperforming XGBoost and LSTM in most configurations.
  • Segment Length: A segment length of 50 time steps yielded the best performance, particularly when combined with shapelet-based feature extraction.
  • Ensemble Combination: Combining predictions from different segment lengths (50, 1000, 2000) via a logical OR operator improved performance, with the combination of 50 and 1000 showing the strongest results on the private leaderboard.
• Efficiency: The experiments were conducted on a standard Mac Mini (M1 chip), with an average computation time of 10 hours per experiment, indicating the method does not require specialized high-performance computing hardware for training.

Significance
The paper claims that the hierarchical modeling approach is highly effective for detecting subtle anomalies in realistic satellite telemetry. The significance lies in the successful integration of shapelet mining with statistical features within a strict anti-leakage framework. The results demonstrate that even with a relatively simple cross-channel aggregator (Logistic Regression), a well-structured ensemble that respects data independence and leverages diverse feature representations (statistical and shapelet) can achieve top-tier performance in a complex, high-dimensional space domain challenge. The work validates that prioritizing robustness, precision, and interpretability through hierarchical ensembling is a viable strategy for operational satellite monitoring.