EitaNis/pdm-synth-compressor-bearing

PdM Synthetic Compressor Bearing-Fault Dataset

Physics-composed synthetic data for compressor bearing-fault detection, generated by the synth package of the PdM predictive-maintenance system. 2000 randomised compressor instances are each simulated over a degradation trajectory, producing both an ISO 10816-3 vibration-velocity RMS feature with fully consistent derived labels, and panel-level aggregate current waveforms composed by summing the compressor's own motor current with a background-load model — the (aggregate, per-load ground truth) pairs a NILM disaggregation model trains on.

This is the pretraining half of a sim-to-real study; it is published so the transfer claim can be checked against the exact bytes the model saw, rather than against a regeneration of them.

Provenance and licence

• Original source: https://github.com/Lior-Nis/pulse
• Licence: cc-by-4.0
• Attribution required: PdM project, synth package. Released under CC BY 4.0.

Why this redistribution is permitted

This dataset is generated in full by code in this project's synth package; it contains no third-party data and no measurements from any external source. It is released by the authors under CC BY 4.0.

Configurations

• nilm — splits: test, train, validation
• scalar — splits: test, train, validation

from datasets import load_dataset

ds = load_dataset("EitaNis/pdm-synth-compressor-bearing", "nilm")
ds = load_dataset("EitaNis/pdm-synth-compressor-bearing", "scalar")

How the signals are generated

Each instance draws randomised motor and bearing parameters (rated power, shaft speed, ball count, pitch and ball diameter, contact angle) and a randomised degradation trajectory (onset day, duration, accelerating-growth exponent). One canonical scalar severity state 0->1 drives both channels, so the vibration and current signals are physically consistent with each other rather than independently curve-fitted.

• Vibration: broadband velocity RMS in mm/s, judged against the ISO 10816-3 Group 2 zone boundaries (4.5 / 7.1 mm/s) the product itself uses.
• Current: a 50 Hz fundamental (Israel grid) with MCSA sidebands at f_line +/- BPFO, sideband amplitude scaled by severity. Outer-race defects (BPFO) only.
• Aggregate: the compressor's current summed with a background-load model carrying a business-hours duty cycle, percent-level odd harmonics of the line frequency (nonlinear loads), and broadband noise.

Labels

All label columns are views of the one severity trajectory, so they are mutually consistent by construction:

• severity — the canonical continuous ground-truth state, 0->1.
• fault_type — healthy while severity is 0, else bearing_wear.
• threshold_state — healthy/warning/critical, from applying the ISO thresholds to the noisy observed RMS, so it matches what a deterministic threshold engine would compute from the same sensor reading.
• rul_days — days until the noiseless trajectory crosses critical, solved in closed form. Null when the crossing is already past, and null for all time on instances that plateau below critical; reaches_critical distinguishes those two cases and is constant per instance.

Splits

Grouped by instance, never by row: an instance's rows at different days share a bearing, a defect and a noise seed, so a row-level shuffle would leak. 70% of instances train, 10% validation, 20% test, assigned by instance index so the partition is reproducible without shipping an index file.

Composition

• 26,000 rows in scalar, 5,200 in nilm, over 2,000 instances.
• 50.4% of scalar rows carry a bearing defect; 13.6% are at the critical threshold.
• 99.9% of nilm rows past severity 0.5 have a sideband recoverable from the aggregate above the harness's SNR threshold (separable).
• In nilm, background_current is not stored; it is exactly aggregate_current - compressor_current.

Limitations

This is simulated data and the sim-to-real gap is unresolved — that is what the companion real-data benchmark exists to measure. Specifically: the signal model is parametric (analytic sidebands and harmonics), not an electromagnetic simulation of an induction motor; the MCSA sideband-to-fundamental ratio is order-of-magnitude plausible rather than calibrated against measurements; only outer-race defects are modelled; the background-load model is a stand-in, not a per-appliance panel model; and the line frequency is 50 Hz throughout, so sideband positions differ from 60 Hz-grid data.

Citation

@software{pdm_synth_compressor_bearing,
  title  = {PdM Synthetic Compressor Bearing-Fault Dataset},
  author = {PdM project},
  year   = {2026},
  note   = {Generated by the synth package; physics-composed NILM/MCSA data}
}