Skip to content

SKAB (Skoltech Anomaly Benchmark)

What is this data? A physical water-circulation testbed at Skoltech: a pump, valves and 8 sensors (vibration, pressure, current, flow, temperature). Each of the 35 short experiments injects a real physical fault (valve closing, cavitation, ...). A rare benchmark with physically induced, precisely timed faults.

  • Source: https://github.com/waico/SKAB (Katser & Kozitsin, 2020)
  • Contents: water-circulation testbed with 8 sensors; 35 experiments (valve1/valve2/other) plus one anomaly-free run, 1-second sampling
  • License: the repository is AGPL-3.0 — code is never copied. Only the data CSVs are used (the authors state the data is free to use; citation required).
  • Download: tsad-forge download skab or --subset valve1
  • Loader: load_skab(experiment="valve1/0") — train comes from anomaly-free.csv

Known flaws

  • Short experiments (~1k steps each) — too little data for most deep models.
  • Separate changepoint column: this loader uses only the anomaly labels; changepoints are ignored.
  • Operating conditions differ between the anomaly-free (train) run and the test experiments — a covariate shift exists by construction.

Quick EDA (from local data)

Sample test channels with labeled anomalies shaded, plus the event-length distribution. Regenerate with tsad-forge viz after downloading the data.