deepla

Turn event logs into a live risk signal.

deepla models industrial event chains (timestamp · machine · alarm · severity) as a language of events, detects anomalies, recalibrates mis-tuned alarms, and surfaces a single 0–100% calibrated probability of risk.

Try the demo → How it works
0% probability of risk

Live gauge — value animates as the event stream evolves.

0–100% Calibrated risk probability
3 Public benchmarks · SKAB, SMD, SMAP/MSL
26 Tests passing
local-first No cloud — data never leaves the machine

Built for noisy, real-world logs

Event-chain modeling

Multivariate Hawkes + sequence models treat a log as a language of events — capturing bursts and cross-machine excitation.

Alarm recalibration

Stores alarm value, threshold and ground truth to measure and fix mis-calibrated alarms per machine.

Honest validation

Reports FDR, detection delay and alarm volume — and publishes the negative results too, not just the wins.

One risk number

P(≥1 high-severity event in the horizon), calibrated and shown as a single 0–100% gauge.

From alarm flood to one number

Before 
22:04:01  ALARM  HIGH      M003
22:04:01  ALARM  HIGH      M003
22:04:02  ALARM  MEDIUM    M001
22:04:03  ALARM  HIGH      M005
22:04:07  ALARM  CRITICAL  M003
22:04:09  ALARM  MEDIUM    M002
22:04:11  ALARM  HIGH      M003
22:04:12  ALARM  LOW       M006
…         235,118 more events

Every alarm looks equal. No priority, no horizon, no calibration.

After
68%

One calibrated risk probability per machine over the next horizon — with FDR, detection delay and alarm volume measured at a matched alarm budget.

How it works

  1. 1 Event generator → synthetic logs with labeled ground truth
  2. 2 Storage → append-only time-series (SQLite today, Timescale next)
  3. 3 Risk engine → calibrated p_risk ∈ [0,1]
  4. 4 Dashboard → gauge + time series + event feed

Run it locally

make gen        # synthetic event stream + labelled ground truth
make train      # calibrate the risk engine
make dashboard  # 0-100% gauge + time series + event feed

Local-first: SQLite + a NumPy risk engine, no external services.

About

deepla is a research-grade prototype: spec-driven, local-first, and validated before deployment. deepla.org

Live demo

Explore the risk dashboard on real synthetic data — gauge, time series and event feed, computed by the actual engine. Open demo.deepla.org

Fair-use policy. The demo runs on shared resources: sessions are time-limited and the number of concurrent users is capped. It uses synthetic data only; no real plant data is exposed.

Get in touch

Or email us directly: [email protected]