smilegeng/industrial-sensor-anomaly-data

Industrial Equipment Sensor Anomaly Data (TsFile)

This dataset is a lossless conversion to the Apache TsFile format of the HuggingFace dataset Petsteb/industrial-sensor-anomaly-data, a synthetic multivariate sensor benchmark for anomaly detection.

Original dataset

• Source dataset: Petsteb/industrial-sensor-anomaly-data
• License: MIT
• Content: fully synthetic sensor data from a simulated manufacturing plant with 5 equipment units (EQ-001 .. EQ-005). Each unit produces 10,000 one-minute readings across 11 sensor channels plus metadata, derived features, and operating-mode labels. ~4.5% of readings are anomalies across 4 anomaly types (thermal runaway, bearing degradation, pressure leak, sensor malfunction). Reproducible with numpy.random.default_rng(seed=42).

Scale

• 50,000 rows = 5 equipment units × 10,000 one-minute readings each
• 22 source columns → 21 stored (see Dropped column below)
• Time range: 2024-01-01 00:00 → 2024-01-07 22:39 (1-minute cadence)
• The 5 units share one time axis; (equipment_id, timestamp) is unique.

TsFile storage mapping (table model)

Role	Column(s)	Type	Notes
TAG	equipment_id	STRING	EQ-001 .. EQ-005; one unit = one device
Time	source timestamp	INT64 (ms)	per-minute timestamp, time primary key
FIELD	temperature_c, vibration_mm_s, pressure_kpa, motor_rpm, flow_rate_lpm, power_consumption_kw, coolant_temp_c, acoustic_level_db, oil_viscosity_cst, humidity_pct, ambient_temp_c	DOUBLE	11 sensor channels (the first 10 carry intentional NaN gaps)
FIELD	equipment_age_hours, hours_since_maintenance, rolling_anomaly_rate, maintenance_priority_score	DOUBLE	derived / metadata features
FIELD	is_anomaly	INT64	0 / 1 anomaly label
FIELD	operating_mode, anomaly_type, alert_code	STRING	categorical labels

Conversion notes

• TAG = equipment_id (5 devices). Required so that each device's time axis is strictly monotonic — the same timestamp occurs once per unit.
• Time: source timestamp parsed to INT64 epoch milliseconds; the original text column is dropped (its information is preserved losslessly in Time). Rows are sorted ascending by (equipment_id, Time).
• Dropped column (with consent): reading_id (a global row-id surrogate key R000000..R049999) is removed. Once Time + equipment_id identify a row it is a redundant key with no time-series signal. This is the only column dropped.
• Nulls kept as-is: the 11 sensor channels contain intentional NaN gaps (~600–1300 per column) from the synthetic generator. They are preserved — TsFile simply does not write null cells. No rows were dropped (50,000 in, 50,000 out; per-unit null counts and the is_anomaly distribution match the source exactly).
• Single file: 50,000 rows is below the tool's 2²⁰ = 1,048,576-row shard threshold, so the output is one .tsfile.

Layout

data/
└── industrial_sensor_anomaly.tsfile

Usage

from tsfile import TsFileReader

reader = TsFileReader("data/industrial_sensor_anomaly.tsfile")
schemas = reader.get_all_table_schemas()
tname = next(iter(schemas))

cols = ["equipment_id", "temperature_c", "vibration_mm_s", "is_anomaly", "anomaly_type"]
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{industrial_sensor_anomaly_data,
  title     = {Industrial Equipment Sensor Anomaly Data},
  author    = {Petsteb},
  year      = {2025},
  url        = {https://huggingface.co/datasets/Petsteb/industrial-sensor-anomaly-data},
  publisher = {Hugging Face}
}

Original dataset licensed under MIT.