Satellite Telemetry Anomaly Detection

Download the NASA SMAP and MSL anomaly detection datasets from the Numenta Anomaly Benchmark or the original NASA JPL release. These contain real telemetry channels from the Soil Moisture Active Passive satellite and Mars Science Laboratory rover, with labeled anomaly windows.

Load the data using pandas and plot representative channels. Understand the structure: each channel is a univariate time series, but anomalies often manifest across multiple correlated channels simultaneously — which is why multivariate methods outperform per-channel thresholds.

Raw time-series values alone are weak features. Engineer richer representations: rolling statistics (mean, std, min, max over windows of 10, 50, 200 samples), rate of change (first and second derivatives), frequency-domain features (FFT power in key bands), and cross-channel correlations.

Normalizing features is critical here — telemetry channels span wildly different value ranges. Use RobustScaler from scikit-learn, which is resistant to the outliers you're trying to detect.

Before applying ML, implement a simple baseline: flag any reading more than 3 standard deviations from the rolling mean as an anomaly. Calculate precision and recall on the labeled dataset. This baseline is what every ML method must beat to justify its complexity.

You'll find that thresholds miss slow-drift anomalies entirely and generate excessive false alarms during legitimate operational mode changes — exactly the weaknesses that motivate ML approaches.

Train an Isolation Forest on the feature-engineered training windows (normal data only). Isolation Forest explicitly models anomalies as points that are easy to isolate — requiring fewer splits in random trees. Apply to the test set and tune the contamination parameter.

Also apply DBSCAN clustering: points that don't belong to any dense cluster are flagged as anomalies. DBSCAN requires no labeled data at all and naturally handles multi-modal nominal behavior (different operational modes cluster differently).

Implement an LSTM Autoencoder in PyTorch or Keras: the encoder compresses the input window into a latent vector, and the decoder reconstructs it. Train only on normal data. At inference time, high reconstruction error signals anomalous behavior that the model never learned to compress well.

This approach captures temporal dependencies that Isolation Forest misses. The reconstruction error time series itself becomes a rich signal — plot it alongside the raw telemetry to develop intuition.

Compute precision, recall, F1, and Area Under the ROC Curve (AUC-ROC) for all three methods (threshold, Isolation Forest, LSTM Autoencoder) on the labeled test set. Plot ROC curves on a single figure for comparison.

Critically analyze failure modes: when does each method miss anomalies (false negatives)? When does it fire false alarms? For satellite operations, false negatives that miss real failures are far more costly than false alarms — discuss how to tune the precision/recall tradeoff accordingly.