Industrial-MultiSensor-AnomalyTS

Synthetic, physics-informed 5-channel condition-monitoring time series from a 40-asset rotating-machinery fleet, with 4 injected fault archetypes; companion LSTM-Autoencoder reaches point-adjusted F1 ~0.9 on held-out (asset-grouped) test assets.

About Dataset

Industrial-MultiSensor-AnomalyTS simulates minute-resolution condition-monitoring telemetry from 40 rotating-machinery assets (pumps, compressors, induction motors), each with 5 sensor channels: vibration RMS velocity, bearing housing temperature, motor current draw, shaft rotational speed, and acoustic emission level. Each channel combines a non-stationary trend, seasonal component, AR(1)-correlated noise, and light cross-channel coupling (current and vibration co-vary with shaft speed). Four fault archetypes are injected as labeled anomaly windows: gradual bearing wear, rotor imbalance, electrical fault, and sensor dropout, drawn from the rotating-machinery diagnostics literature. The bearing wear archetype deliberately delays its label onset until the degradation ramp passes 55% severity, leaving a genuine pre-onset early-warning window rather than only a nowcasting target. The dataset is fully synthetic, containing no proprietary or real-plant data. Baseline sensor ranges are illustrative order-of-magnitude values consistent with published vibration-severity charts (ISO 10816/20816) and typical induction-motor operating temperatures, not fit to any specific real asset. The train/val/test split is grouped by asset ID, so evaluation measures generalization to unseen machinery rather than memorization of one asset's noise pattern.

Purpose of Dataset

This Dataset Supports Research On Predictive Maintenance And Anomaly Detection For Industrial Rotating Machinery, Particularly Reconstruction-based Unsupervised Methods That Must Generalize Across A Fleet Of Assets Rather Than Being Tuned To One Specific Machine. It Targets Researchers And Engineers Building Condition-monitoring Systems That Need To Distinguish Genuine Multi-channel Correlated Faults From Routine Sensor Noise, And Evaluates Whether Such Systems Can Detect Degradation Before A Fault Is Fully Developed, Not Only After The Fact. The Asset-grouped Evaluation Design Directly Tests Generalization To Unseen Equipment, Which Matters For Deployment Across Heterogeneous Industrial Fleets.