AI4Manufacturing/MENDELEY-rotor-perception

Mendeley rotor faults — perception representations (visual grounding)

Part of the AI4Manufacturing FORGE corpus (Category C, task T-C1) and its first rotor-family dataset: the fault is in the shaft line — unbalance, misalignment, mechanical looseness — not in a bearing or a gear. One-second vibration windows from a belt-driven rotor bench, rendered as perception images, one HF config per representation.

Records: 2000 across 4 configs (500 windows each); labels {'normal': 125, 'misalignment': 125, 'unbalance': 125, 'looseness': 125}.

Configs

load_dataset("AI4Manufacturing/MENDELEY-rotor-perception", "spectrogram")

config	records	splits
spectrogram	500	{'train': 400, 'test': 100}
scalogram	500	{'train': 400, 'test': 100}
waveform	500	{'train': 400, 'test': 100}
reshaped	500	{'train': 400, 'test': 100}

Schema (7-field unified record)

field	meaning
query	the classification instruction (one of 30 deterministic paraphrases per representation)
image	the rendered signal image (bytes embedded)
annot	gold rotor condition: normal / misalignment / unbalance / looseness
reasoning	empty — this release has no reasoning track (see Why perception-only)
cate / task	C / T-C1 (signal fault classification)
metadata	JSON string: representation, features (time-domain stats), computed_verdict, computed_indication, harmonic_profile, harmonic_elevation, evidence_tier, evidence_is_gate: false, channel, fr_hz + fr_source, test_id, window_idx, fs, image_sha256, split

Splits

train / test, test-wise (leakage-safe): each of the 20 recorded tests is a distinct physical assembly and lives wholly on one side; the last test of each condition is held out. Windows never cross tests.

Why perception-only (measured, not assumed)

There is deliberately no reasoning/CoT sibling. A label-free shaft-order detector (forge_tools.shaft_harmonics) was run against a same-rig healthy baseline; its per-class firing rate and median 1× elevation, measured at build time, are:

gold class	detector fires rotor_anomaly	median 1× elevation	1× elevation range
normal	0.088	1.6	0.3 – 3.9
misalignment	0.008	1.4	0.5 – 2.5
unbalance	1.000	28.2	18.2 – 38.9
looseness	0.944	13.4	1.6 – 28.1

Read that table honestly: unbalance and looseness are strongly detectable, but misalignment is indistinguishable from healthy — its entire 1× elevation range sits inside the healthy range, so no threshold can separate them. That is the physics, not a bug — a misalignment signature is primarily axial, and this bench has four radial accelerometers and no axial one. Unbalance and looseness are also not separable from each other here (both simply lift 1×). And an absolute harmonic gate is useless on this rig: its belt/pulley drivetrain means the healthy state already carries a strong 2×/3× ladder (its healthy 2× sits ~9× above its own 1×), so 'harmonic structure present ⇒ fault' fires on 97.6% of healthy windows (122/125 in this build).

Because a compute-then-check chain-of-thought could not honestly reach 2 of the 4 classes, the labels ship as implanted gold (the rig operator's documented condition) and reasoning stays empty. The shaft-order evidence still travels on every record as informational metadata (evidence_is_gate: false) — nothing was dropped, relabelled or filtered by it, so you can audit the claim above yourself from the published rows.

Provenance & reproducibility

Generated deterministically by forge_agent/examples/mendeley_rotor/convert.py (e6069b7005) → forge_model/MENDELEY_ROTOR/convert_mendeley_rotor.py (7708e1353d); see provenance.json.

25 evenly-strided 1 s windows per test (consecutive seconds are highly correlated; striding avoids near-duplicate rows), channel 0 (V_coupling), shaft rate estimated per test with a bounded+gated search that falls back to the documented nominal and records fr_source. Images carry no titles and no condition text in their filenames (answer-leak hygiene).

Caveats

• Not a reasoning dataset. Use it for representation diversity / visual grounding.
• Misalignment is unlearnable-from-physics here and may be hard for a model too; the evidence table above is the honest prior.
• One bench, one speed regime (~21.8 Hz rotor, belt-driven). Cross-rig generalization is untested.
• Radial channels only — no axial instrumentation.

Source & license

Source: Mechanical faults in rotating machinery dataset (normal, unbalance, misalignment, looseness) — L. Brito, G. A. Susto, J. N. Brito, M. Duarte, Mendeley Data, doi:10.17632/zx8pfhdtnb.1. License: CC BY 4.0; this derived dataset is redistributed under the same terms with attribution to the original authors.