AI4Manufacturing/OTTAWA-VARSPEED

Ottawa variable-speed bearing — race damage from the order spectrum (reasoning track)

Part of the AI4Manufacturing FORGE corpus (Category C, task T-C1), and the corpus's first dataset recorded under changing shaft speed. Each record is the order spectrum of one 10 s vibration record's envelope, with the outer-race (BPFO) and inner-race (BPFI) orders marked. reasoning is empty here; the OTTAWA-VARSPEED-annotated sibling fills it.

Records: 35, shipped unsplit (see Splits below); labels {'combined': 11, 'healthy': 12, 'inner_race': 12}; evidence tiers {'weak': 11, 'confirmed': 24}.

Why the horizontal axis is orders, not hertz

A bearing fault frequency is a fixed multiple of shaft speed. Hold the speed constant and it lands on one line in hertz; sweep the speed and the same defect paints a smear across the spectrum, because it was at 90 Hz at the start of the record and 150 Hz at the end.

So the abscissa here is shaft order — events per shaft revolution — not hertz. The pipeline is: pick the demodulation band with a spectral-kurtosis kurtogram, band-pass and Hilbert-envelope in the time domain (the carrier band is a range in hertz, so this step has to happen before any resampling), then resample the envelope onto shaft angle using the encoder, then transform. On an order axis the outer-race and inner-race lines sit still no matter what the speed does.

Rig

University of Ottawa variable-speed bearing rig: an ER16K test bearing on a shaft driven through a range of speeds, with an accelerometer on the housing and a 1024 pulse/rev incremental encoder on the shaft, both sampled at 200 kHz for 10 s. Sixty records: five bearing states × four speed profiles × three repeats.

speed profile	what the shaft does	records
rising	speeds up throughout	15
falling	slows down throughout	15
rise_then_fall	speeds up, then slows	15
fall_then_rise	slows, then speeds up	15

The shaft rate averages 16.8–22.4 Hz across records and moves by roughly a factor of two within each one. Nothing here is a steady-state measurement.

⚠️ Three of the five classes are here; two are not

healthy, inner_race and combined reach this repo. outer_race and ball do not — no line clears the magnitude screen on any of their 24 records, so there is nothing for a chain-of-thought to cite. They ship in AI4Manufacturing/OTTAWA-VARSPEED-perception instead. Every query still lists all five classes, so a model must rule the absent two out rather than be told they are impossible; but you cannot measure five-way accuracy on this repo, and a model trained only here will never have seen an outer_race or ball record.

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 class: healthy / inner_race / outer_race / ball / combined
reasoning	chain-of-thought (empty here; filled in the -annotated sibling)
cate / task	C / T-C1 (signal fault classification)
metadata	JSON string: unit, sweep, repeat, fs, fr_hz_mean, cpr, samples_per_rev, abscissa, demod_band_hz, speed_varies, bearing_wise_split, split_note, calibration_side, multi_label_source, upstream_damage_description, representation, track, track_reason, computed_verdict, computed_snr, evidence_tier, evidence_measured_on, baseline_used, baseline_drawn, image_sha256

Splits

This dataset is shipped unsplit, deliberately — there is no defensible train/test division to give you.

The source uses one physical test bearing, swapped between damage states between runs. Every record therefore shares the same bearing, the same housing, the same sensor mount and the same shaft. There is no held-out unit to move to a test set, and no grouping of these 60 records puts genuinely unseen hardware on the far side. Publishing our own division would hand you one particular choice as if it were the only defensible one; the honest thing to report is that none of them is.

Note on the train split name. HuggingFace requires every split to be named, so the single unsplit set is served as train. It is the whole dataset, not a training portion — there is no matching test. It returns all 35 records, and dividing them is on you.

What we did use, and why it is metadata rather than a split. The healthy baseline and the magnitude screen were fixed on the two monotonic speed profiles (rising, falling) and then applied unchanged to the two reversing ones (rise_then_fall, fall_then_rise). metadata.calibration_side records which side each record is on ({'calibrate': 18, 'blind': 17}). That is evidence the screen was not tuned per record — it is not a leakage-safe holdout, because the same bearing in the same damage state appears on both sides. Do not train on one side and report on the other as if it were a test set.

If you need a division, metadata.sweep and metadata.repeat are the two axes available. Whatever you pick, say in your paper that it is not bearing-wise.

Provenance & the evidence gate

Generated deterministically by forge_agent/examples/ottawa_varspeed/convert.py (df86f4a36e) → forge_model/OTTAWA_VARSPEED/convert_ottawa_varspeed.py (2eff095405); full record in provenance.json.

Labels come from the source's own file naming. An evidence gate — a label-independent computation on the raw vibration — decides what may enter the reasoning track: a line is counted as present when it clears a signal-to-noise floor of 2.5 and stands 3.0× above the healthy baseline at the same order. The search never sees the label.

Both numbers were restandardised on this rig, from the calibrate-side healthy records only; they are not carried over from the constant-speed Ottawa set or from any other dataset in the corpus. The kurtogram is capped at 15.0 kHz and the analysis runs to order 12.0.

This track keeps a record when the blind verdict is compatible with the gold: identical for healthy and inner_race, and either race for combined, which is genuinely both. 1 of the 36 records routed here (a combined one) was dropped for showing nothing at all.

Caveats

• ⚠️ A bearing-wise split is impossible on this rig, which is why none is shipped. The source uses one physical test bearing, swapped between damage states between runs. Every record therefore shares the same housing, the same sensor mount and the same shaft. There is no held-out unit to move to a test set, and metadata.bearing_wise_split is false on every row so the constraint travels with the data. A model can score here by recognising the rig; treat every number measured on this dataset as an upper bound and confirm transfer on a dataset with more than one bearing.
• outer_race and ball ship perception-only (24 records). No line clears the magnitude screen on any of their records — not a majority, all of them. The screen is not simply deaf: on the same threshold, healthy reads healthy on 12 of 12 (no false positives) and inner_race reads inner-race on 12 of 12. Whatever the reason, those two classes carry no citable evidence here, so their order spectra are published nowhere and only their texture views ship.
• combined is multi-label at the source but ships as one string. All 12 combined records carry inner- and outer-race damage on the same bearing; annot is the single token combined and metadata.multi_label_source is true (11 of them are in this repo). The blind search finds the inner-race line on 11 of 12 and the outer-race line on none of them, so the label is right while the visible evidence is only half of it. Do not read a combined record as showing two lines.
• The source describes the damage for only three of the five classes. It names the fault in the filename but documents no defect size, origin or method of introduction for ball or combined — 24 records dataset-wide, 11 of them in this repo, carrying metadata.upstream_damage_description: false. This is the same class of gap that made the PHM 2010 wear labels unusable at face value, so it is recorded rather than glossed.
• The evidence gate is measured on the SIGNAL, not on the image. The line search reads the resampled envelope; it never opens the PNG. A confirmed tier says this signal supports the label, not this image shows it. Every record carries evidence_measured_on: "signal".
• 60 records is small. One rig, one bearing, five states, four speed profiles, three repeats. The value here is the kind of evidence — a fault line that holds still while the speed moves — not the volume.

Source & license

Source: University of Ottawa bearing vibration dataset under time-varying rotational speed. Please cite: Huang & Baddour, Bearing vibration data collected under time-varying rotational speed conditions, Data in Brief 21 (2018) 1745–1749, doi:10.1016/j.dib.2018.11.019. Data: doi:10.17632/v43hmbwxpm.2, CC BY 4.0.