Telemetry Anomaly Detection

Real-time anomaly detection for satellite and industrial sensor telemetry. Point it at a stream of sensor readings and it surfaces the moments that matter — degradations, faults, off-nominal events — each with a confidence score and the detectors that fired. No labelled training data required.

Screenshots

Authenticated ops console — live telemetry, anomaly monitor, sample-dataset detection.

Why

Satellite and plant operators spend hundreds of engineer-hours a month eyeballing telemetry dashboards. Real degradations accumulate quietly while on-call engineers get paged at 3 AM for perfectly normal orbital patterns. A missed anomaly in orbit can cost anywhere from $50K to an entire mission. The goal here is simple: catch the real events automatically, and stay quiet on everything else.

How it works

Telemetry (JSON / CSV / InfluxDB) → feature engine → 6-detector ensemble → consensus + root cause → alert

• A consensus ensemble of six detectors — classical statistical methods alongside deep sequence models (LSTM and TCN, trained per channel) — votes on every window. An event is only raised when detectors agree, which is what keeps the false-positive rate low.
• Per-sensor models ship in SKAB-S16/ (Accelerometer, Current, Pressure, Temperature, Thermocouple, Voltage, Volume Flow Rate).
• Every alert carries a confidence score and a root-cause hint — not just a red flag.

Results

Validated blind (no prior knowledge of the events) against real telemetry:

• ISS (NORAD 25544): 4 of 4 known events detected on the validation set, with zero false positives on normal orbital operations — cross-checked against public NASA/AMSAT records.
• ESA Mission 1: 58 channels, ~7.1M telemetry points processed.
• Also benchmarked on the SKAB industrial sensor dataset (trained models included in the repo).

Full methodology and tables are in docs/BENCHMARK_RESULTS.md.

Run it

cp .env.example .env     # configure inputs and alerting
docker compose up        # starts the API + detectors

A static dashboard for exploring detections lives in dashboard/ — open dashboard/index.html and load one of the example feeds in dashboard/samples/ (satellite thermal, spacecraft EPS, valve sensor).

Stack

Python · PyTorch (LSTM, TCN) · NumPy / Pandas · Docker & docker-compose · AWS

Status

A working detection engine with reproducible benchmarks and a demo dashboard. Input connectors and alert routing are configurable; see docs/ for the detection principles, FMEA mapping, and the broader design notes.