한국어

BIoTan

Zero-config, peer-relative anomaly detection for fleets of homogeneous IoT assets.

BIoTan looks at a fleet of similar devices — solar inverters, refrigeration units, pumps, drives, sensors — and tells you which ones are drifting away from their peers, without you configuring thresholds or manually grouping anything.

It does this by comparing each device not against an absolute limit, but against what its peers are doing at the same moment. Shared conditions — weather, load, seasonality — affect every peer equally and cancel out, leaving only the deviations that actually matter.

This repository is the free, open core: a batch backtesting engine. You give it historical sensor data as CSV; it gives you back cohorts, per-device deviation timelines, and flagged assets with reasons.

Why peer-relative?

Most monitoring asks "is this value above a threshold?" — which means someone has to set, tune, and maintain that threshold for every device, and it fires false alarms whenever a shared condition (a cloudy day, a heavy-load shift) moves the whole fleet at once.

BIoTan asks a different question: "is this device behaving differently from its peers right now?" That requires no per-device setup, automatically ignores fleet-wide common-mode changes, and surfaces the genuinely odd unit — the inverter that's 12% below its neighbors after sun-angle correction, the drive accumulating bad sectors faster than its cohort, the engine drifting from the healthy baseline as it nears failure.

What it does

• Auto-clustering — discovers behavioral cohorts from the data. No manual tagging.
• Common-mode removal — compares each device to its cohort peers at each timestamp, using robust statistics (median / MAD) so a few failing peers don't poison the baseline.
• Multi-signal detection — different faults look different, so it tracks several orthogonal signals: persistent offset, gradual change/drift, instability, and rigidity (a sensor that varies far less than its peers — i.e. stuck).
• Effect-size gating — flags require both statistical and practical significance, so near-zero noise doesn't trigger alerts.
• Backtest timeline — if you provide known failure/replacement dates, it shows when a device first started diverging — i.e. how many days earlier you could have known.

What it does not do (and why)

Being honest about the boundary is the point.

• It is not a guaranteed failure predictor. Some failures happen with no prior signal in the data; no peer-relative method can catch those. BIoTan is a risk-prioritization and degradation-tracking tool — it tells you what to look at first, not that everything unflagged is safe.
• It does not catch contextual anomalies well — cases where a value is within normal range but wrong for its context/timing. Those need richer time-series models or user labels.
• Backtest lead-times are an optimistic upper bound. Because backtesting tunes to data that already happened, real-time results may differ.
• This core is batch-only. Real-time ingestion, MQTT/stream/database connectors, fleet operations, alert delivery (Slack/PagerDuty/email), and multi-node management are not part of this repository.

Validation

The core method has been tested across seven independent datasets spanning very different sensor physics — synthetic fleets, real air-quality and climate data, real hard-drive SMART telemetry with true failure labels, NASA turbofan degradation, and real satellite telemetry. The same pattern held throughout: where a location/condition leaves a strong enough signature, peer-relative deviation tracks real problems; where the signal is weak or the anomaly is contextual, the simple method reaches its limit.

A reproducible validation against real public data is included in /validation — run python validation/run_cmapss.py to download the NASA C-MAPSS FD001 turbofan fleet (100 engines, run to failure with true failure points) and reproduce the numbers there. On that data, common-mode removal lifts peer-z above 2σ before failure for 99/100 engines, and the conservative zero-config gate confirms the fastest-degrading engines with a median lead of ~11 cycles. It also honestly surfaces a limit: the behavioral-profile clustering assumes daily-cyclic data, so a non-cyclic run-to-failure fleet is best analysed as a single cohort (the script shows both).

Quick start

pip install -r requirements.txt
python -m biotan backtest --input your_data.csv --out report.html

Your CSV needs at least three columns: device_id, timestamp, value. Optional: metric, group, unit. Everything runs locally — no data ever leaves your machine, and there is no telemetry.

Try it on synthetic data with known cohorts and injected faults:

python scripts/make_synthetic.py --out demo.csv --faults 2 --validate
python -m biotan backtest --input demo.csv --labels demo.faults.csv --out report.html

Commands

The engine runs as one batch pipeline, but each stage is also runnable on its own (all read a CSV; nothing is configured by hand):

command	what it does
python -m biotan summarize --input data.csv	parse + normalize; print fleet summary and inferred cadence
python -m biotan cluster --input data.csv	discover behavioral cohorts (auto, zero-config)
python -m biotan peerz --input data.csv	peer-relative deviation (peer-z) timelines, common mode removed
python -m biotan signals --input data.csv	the four detection signals per device
python -m biotan flag --input data.csv	apply the effect-size gate; list flagged devices + reasons
python -m biotan backtest --input data.csv --out report.html [--labels failures.csv]	full pipeline → self-contained HTML report (+ lead time if labels given)

A labels CSV (for lead time) needs device_id and fault_start (optional metric). The HTML report is a single self-contained file — inline SVG charts, no external assets, no network calls.

License

BIoTan-core is source-available under the PolyForm Noncommercial License 1.0.0. Free for evaluation, research, and noncommercial use. Commercial or production use, and offering it as a hosted service, require a separate commercial license — contact contact@frontli.ne.kr.

Copyright (c) 2026 Victor Minbeom Joo d/b/a Front-Line.

---

BIoTan is open core. Connectors, real-time operation, and fleet management are available separately.