Telemetry Anomaly Detection Toolkit

telemetry-anomdet is an open-source anomaly detection toolkit for spacecraft telemetry. It ingests raw telemetry (SMAP, CSV), preprocesses it, and runs a stacking ensemble of classical and deep learning detectors with per-channel SHAP attribution and LLM-generated diagnostic reports, designed to produce actionable diagnostics within the ground station inter-pass window.

Validated on SMAP (NASA), with OPS-SAT (ESA) as a cross-dataset generalization source.

Current features:

  • SMAP and CSV ingestion into long-form TelemetryDataset (load_smap, load_smap_labels, load_from_csv)

  • Preprocessing pipeline: clean, dedupe, resample, interpolate gaps, normalize

  • Windowed feature extraction: statistical features and raw 3D tensors for sequence models

  • BaseDetector interface: unified fit / decision_function / predict / is_anomaly API shared by all detectors

  • PCAAnomaly and KMeansAnomaly classical detectors (3D input, flatten internally)

  • AnomalyEnsemble: stacking combinator with configurable normalization and combine strategy

  • Per-model score decomposition via score_components() (SHAP hook)

  • Point-adjusted evaluation (point_adjust, point_adjusted_f1, best_point_adjusted_f1) and a reproducible SMAP benchmark (examples/smap_benchmark.py)

Coming in the next few months:

  • IsolationForestAnomaly

  • GDN: graph deviation network for inter-sensor relational anomalies

  • TranAD: transformer-based sequence reconstruction

  • SHAPExplainer: per-channel attribution over score_components()

Coming in by the end of 2026:

  • LLM reasoning layer (Llama 3.1 8B on Jetson Orin via llama.cpp)

  • OPS-SAT cross-dataset generalization evaluation

Contents

User Guide

API Reference