smilegeng/controlled-anomalies-time-series-dataset

CATS — Controlled Anomalies Time Series (TsFile)

This dataset is a lossless conversion to the Apache TsFile format of the HuggingFace dataset patrickfleith/controlled-anomalies-time-series-dataset (CATS), a multivariate time-series anomaly-detection benchmark.

Original dataset

• Source dataset: patrickfleith/controlled-anomalies-time-series-dataset
• Author: Patrick Fleith / Solenix Engineering GmbH
• Cite: Patrick Fleith (2023). Controlled Anomalies Time Series (CATS) Dataset (Version 2). https://doi.org/10.5281/zenodo.8338435
• Content: commands, external stimuli and telemetry of a simulated complex dynamical system with 200 injected anomalies, designed for benchmarking multivariate-time-series anomaly detection.

Scale

• 5,000,000 rows at 1 Hz (1 second per row)
• Time range: 2023-01-01 00:00:00 → 2023-02-27 20:53:19 (UTC), strictly increasing, no gaps or duplicates
• 17 variables + 2 ground-truth label columns
• First 1M rows are nominal; the last 4M contain 200 anomalous segments.

What is in this repository

cats_1.tsfile … cats_5.tsfile   # converted time-series (sharded, see below)
metadata.csv                     # 200 anomalous-segment annotations (copied verbatim)

TsFile storage mapping (table model)

Role	Column(s)	Type	Notes
Time	source timestamp	INT64 (ms)	1 Hz, strictly increasing, globally unique
FIELD	aimp, amud, arnd, asin1, asin2, adbr, adfl	DOUBLE	stimuli / commands
FIELD	bed1, bed2, bfo1, bfo2, bso1, bso2, bso3, ced1, cfo1, cso1	DOUBLE	commands / telemetry
FIELD	y	INT64	anomaly label (0 / 1)
FIELD	category	INT64	anomaly type (0..13)

There is no tag column: CATS is a single system's multivariate series.

Conversion notes

• Time = timestamp → INT64 epoch milliseconds. The series is a single multivariate sequence, strictly increasing and unique, so no tag column is used.
• 17 variables kept as DOUBLE; y and category are integer-valued and kept as INT64 (the anomaly ground truth — the core of this benchmark).
• No columns dropped, no rows dropped: all 5,000,000 rows preserved.
• Sharding: the tool emits one .tsfile per 2²⁰ (1,048,576) rows, so the data is split into 5 shards cats_1.tsfile … cats_5.tsfile; together they form the complete series.
• metadata.csv (the 200 anomalous-segment annotations: start_time, end_time, root_cause, affected, category) is copied verbatim.

Usage

from tsfile import TsFileReader

reader = TsFileReader("cats_1.tsfile")
schemas = reader.get_all_table_schemas()
tname = next(iter(schemas))

cols = ["arnd", "bfo2", "cfo1", "y", "category"]
with reader.query_table(tname, cols, batch_size=65536) as rs:
    while (batch := rs.read_arrow_batch()) is not None:
        df = batch.to_pandas()
        # ... process ...
reader.close()

Citation

@dataset{fleith_cats_2023,
  title     = {Controlled Anomalies Time Series (CATS) Dataset},
  author    = {Fleith, Patrick},
  year      = {2023},
  version   = {2},
  publisher = {Solenix Engineering GmbH},
  doi       = {10.5281/zenodo.8338435},
  url       = {https://huggingface.co/datasets/patrickfleith/controlled-anomalies-time-series-dataset}
}