README.md

metadata

language:
  - en
license: mit
size_categories:
  - 10K<n<100K
task_categories:
  - tabular-classification
tags:
  - anomaly-detection
  - industrial
  - sensor-data
  - synthetic
  - time-series
  - multivariate
pretty_name: Industrial Equipment Sensor Anomaly Data

Industrial Equipment Sensor Anomaly Data

Overview

Synthetic multivariate sensor data from a simulated manufacturing plant with 5 equipment units (EQ-001 through EQ-005). Each unit generates 10,000 one-minute-interval readings across 11 sensor channels, 2 metadata fields, 3 derived features, and equipment operating mode labels.

The dataset is designed for anomaly detection benchmarking. It embeds 4 distinct anomaly types at approximately 4.5% prevalence:

• Thermal runaway — gradual temperature escalation over 15–40 minutes with correlated coolant lag and viscosity drop
• Bearing degradation — vibration spikes with high-frequency noise and proportional acoustic increase
• Pressure leak — slow pressure decay over 20–60 minutes with compensatory flow rate increase
• Sensor malfunction — erratic single-sensor readings that do NOT reflect actual equipment failure (only one sensor affected while correlated sensors remain normal)

Dataset Details

• Rows: 50,000 (10,000 per equipment unit)
• Columns: 22
• Anomaly rate: ~4.5%
• Time span: 1-minute intervals starting 2024-01-01
• Generation: Fully synthetic, reproducible with numpy.random.default_rng(seed=42)

Built-in Complexity

The data contains deliberate challenges for ML pipelines:

• Target leakage features (3): rolling_anomaly_rate, maintenance_priority_score, and alert_code contain information derived from the target label that would not exist at prediction time
• Missing values: ~1–3% per sensor in two patterns — random sporadic dropouts and correlated block outages (5–15 minute windows)
• Concept drift: Baseline sensor values shift partway through the timeline for each equipment unit
• Operating mode effects: Sensor baselines differ significantly across modes (startup, normal, high_load, cooldown, idle)
• Cross-sensor correlations: Temperature affects oil viscosity, vibration drives acoustic levels, RPM and flow determine power consumption

File Structure

• data.csv — Full raw dataset (50,000 rows × 22 columns)

Features

Column	Type	Description
reading_id	string	Unique row identifier (R000000–R049999)
timestamp	datetime	Reading timestamp at 1-minute intervals starting 2024-01-01
equipment_id	string	Equipment unit identifier (EQ-001 through EQ-005)
operating_mode	string	Current mode: startup, normal, high_load, cooldown, idle
temperature_c	float	Main bearing temperature in Celsius
vibration_mm_s	float	Vibration velocity in mm/s
pressure_kpa	float	System pressure in kilopascals
motor_rpm	float	Motor rotational speed
flow_rate_lpm	float	Coolant/fluid flow rate in liters per minute
power_consumption_kw	float	Electrical power draw in kilowatts
coolant_temp_c	float	Coolant outlet temperature in Celsius (lags main temp)
acoustic_level_db	float	Acoustic emission level in decibels
oil_viscosity_cst	float	Lubricant viscosity in centistokes
humidity_pct	float	Ambient humidity percentage
ambient_temp_c	float	Ambient environmental temperature in Celsius
equipment_age_hours	float	Cumulative operating hours of the equipment
hours_since_maintenance	float	Hours elapsed since last maintenance event
is_anomaly	int	Target label: 1 = anomalous, 0 = normal
anomaly_type	string	Anomaly category: normal, thermal_runaway, bearing_degradation, pressure_leak, sensor_malfunction
rolling_anomaly_rate	float	⚠️ LEAKAGE — Rolling mean of is_anomaly with centered window
maintenance_priority_score	float	⚠️ LEAKAGE — Score assigned post-hoc based on anomaly status
alert_code	string	⚠️ LEAKAGE — Alert category derived from anomaly_type with noise

Generation Script

The dataset was generated with the included generate_data.py script using fixed random seeds for full reproducibility. Running the script produces an identical data.csv.

License

MIT — free for any use.