AI4Manufacturing/MAFAULDA-bearing-perception

MaFaulDa — bearing faults, perception representations (visual grounding)

Part of the AI4Manufacturing FORGE corpus (Category C, task T-C1). One-second vibration windows from the seeded rolling-element bearings of the SpectraQuest Machinery Fault Simulator at SMT/COPPE/UFRJ, rendered as perception images — one HF config per representation.

This is the bearing half of the MaFaulDa family; the rotor half ships separately as MAFAULDA-rotor / MAFAULDA-rotor-perception. Same rig, opposite conclusion: the rotor half carries a reasoning track, and this one deliberately does not. Why that is, measured, is the point of this card.

Records: 8960 across 4 configs (2240 windows each); labels {'normal': 98, 'ball_fault': 646, 'cage_fault': 752, 'outer_race': 744}; splits {'train': 1802, 'test': 438}; positions {'none': 98, 'overhang': 1026, 'underhang': 1116}.

Configs

from datasets import load_dataset
ds = load_dataset("AI4Manufacturing/MAFAULDA-bearing-perception", "spectrogram")

config	records	splits
spectrogram	2240	{'train': 1802, 'test': 438}
scalogram	2240	{'train': 1802, 'test': 438}
waveform	2240	{'train': 1802, 'test': 438}
reshaped	2240	{'train': 1802, 'test': 438}

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 bearing condition: normal / ball_fault / cage_fault / outer_race
reasoning	empty on every record — this release has no reasoning track (see Why perception-only)
cate / task	C / T-C1 (signal fault classification)
metadata	JSON string: source_relpath (the unique record key), representation, position, channel + channel_index, fr_hz + fr_source, severity + severity_value, group_id, features, fault_freqs, line_resolvability, demod_band_hz + demod_band_kurtosis, computed_verdict, computed_dominant_line, computed_snr, evidence_tier, evidence_is_gate: false, bearing, file, window_idx, window_start_s, window_seconds, fs, image_sha256, split

source_relpath (e.g. underhang/ball_fault/20g/29.1264.csv) is the unique key. file is only the shaft rate — it is the same string for every position/type/load that was recorded at that speed, so it repeats across hundreds of rows and must not be used as an identifier.

Splits

train / test, severity-wise and leakage-safe: group_id = '<position>|<label>|<severity>', severities sorted numerically (lexically '10g' < '6g', which would hold out the wrong level) and the last one held out. Windows never cross files, and no (position, label, load) assembly spans both sides.

What test actually holds out is an unseen LOAD, not an unseen defect. On the bearing trees the sub-folder is the added imbalance mass (0 g / 6 g / 20 g / 35 g) — an operating condition, not a defect depth. MaFaulDa seeds exactly one physical defect per (position, type), so a defect-instance-wise split is impossible on this dataset. test therefore measures generalization to an unseen load, and a model may still have seen the same physical defect in train. position, severity, severity_value and fr_hz are on every row so a downstream eval can re-carve.

normal/ has no severity level — it is one group (none|normal|none) and sits wholly in train, the same choice CWRU and the rotor sibling make for a single healthy assembly.

Why perception-only — a measured negative result, not an omission

There is deliberately no reasoning/CoT sibling for the bearing half of MaFaulDa, even though the rotor half of the same rig does ship one. Three independent findings, in the order of how much they should convince you.

1. The detector anti-correlates with the gold (measured at build time)

The same label-free envelope chain that works on MFPT and CWRU (bearing_faultfreq → spectral_kurtosis_kurtogram → hilbert_envelope → match_lines, untuned defaults) was run on every window here. Its firing rate is highest on the healthy class:

gold class	windows	fires any fault	verdict == gold	median envelope SNR	max
normal	98	14.3%	85.7%	0.0	32.1
ball_fault	646	2.0%	0.0%	0.0	49.2
cage_fault	752	6.4%	0.0%	0.0	63.0
outer_race	744	3.1%	1.2%	0.0	46.1

Read it plainly: the detector cries fault on 14.3% of genuinely healthy windows and on only 2.0% / 6.4% / 3.1% of real ball / cage / outer-race windows — it is anti-correlated with the truth. verdict == gold is 0.0% / 0.0% / 1.2% on the fault classes; the 85.7% on normal is not a success, it only means the detector found nothing, which is what it also does on most faults. The median winning envelope SNR is 0.0 in every class — against 30–180 on MFPT/CWRU, where this exact chain demonstrably works. The method is not marginal here, it is off by an order of magnitude.

The sharpest single tell: 29 windows receive an inner_race verdict on a dataset that contains no inner-race defect at all. MaFaulDa seeds ball, cage and outer-race faults only, so every one of those is a measured false fire — and it is exactly the fingerprint finding 2 predicts, BPFI landing on the 5× rotor line.

2. The rig's bearing orders coincide with its rotor harmonics

MaFaulDa is a rotor-fault trainer, so every record carries a strong integer shaft-harmonic ladder (1×, 2×, 3×, …). This bearing's defect orders sit on that ladder:

line	shaft order	Δ orders to nearest integer rotor harmonic	resolvable?
BPFO	2.9979×	0.0021	collides
BPFO-1x	1.9979×	0.0021	collides
BPFO+1x	3.9979×	0.0021	collides
BPFOx2	5.9957×	0.0043	collides
BPFOx3	8.9936×	0.0064	collides
BPFI	5.0021×	0.0021	collides
BPFI-1x	4.0021×	0.0021	collides
BPFI+1x	6.0021×	0.0021	collides
BPFIx2	10.0043×	0.0043	collides
BSF	1.8705×	0.1295	escapes
BSF-FTF	1.4957×	0.4957	escapes
BSF+FTF	2.2452×	0.2452	escapes
FTF	0.3747×	0.3747	escapes

BPFO = 2.9979× is 0.0021 orders from 3×; BPFI = 5.0021× is 0.0021 orders from 5×. At the rig's typical ~29 Hz shaft rate that is ≈0.058 Hz of separation, against a 0.500 Hz FFT bin on a 2 s window — 0.12 bins, unresolvable. (This build's 1 s windows give 1.0 Hz bins, i.e. 0.058 bins: worse.) The harmonics land on integers too — 2×BPFO ≈ 6×, 3×BPFO ≈ 9× — and so do the ±1× shaft sidebands, at ≈2×/4× for BPFO and ≈4×/6× for BPFI. Only the cage/ball-spin family escapes: FTF 0.3747×, BSF 1.8705× and BSF∓FTF 1.4957× / 2.2452×. On top of that, match_lines' own ±2% tolerance is ±1.4 Hz at BPFO ≈ 70 Hz, which swallows the 3× rotor harmonic whole — so even a line a finer window could resolve is not separated by the matcher. Every published row carries line_resolvability, marking per line whether it is separated at that record's own shaft rate, so this is auditable record by record.

3. The geometry is correct — this is physics, not a missing constant

The obvious objection is that the bearing constants are simply wrong. They are not: the geometry-derived multipliers reproduce MaFaulDa's own published fault-frequency table to better than 0.0006 orders, and the pair is regression-tested upstream.

line	derived from geometry	MaFaulDa's published table	abs. Δ (orders)
BPFO	2.9979×	2.9980×	0.00014
BPFI	5.0021×	5.0020×	0.00014
BSF	1.8705×	1.8710×	0.00054
FTF	0.3747×	0.3750×	0.00027

(A de-risk pass before the build also swept all six accelerometer axes — underhang/overhang × axial/radial/tangential — on the lines finding 2 leaves standing: fault/healthy envelope-SNR ratios came out 0.40–1.55 with no consistent direction. There is no channel on this rig where the bearing lines lift over healthy.)

What that means for this release

A compute-then-check chain-of-thought over these images could not honestly reach any of the three fault classes, so writing one would be confabulation. Therefore the labels ship as implanted gold — the rig operator's documented seeded condition, which is sound: the defects are physically installed and the runs are steady-state. reasoning is None on every record. No spectrum representation is rendered either — an envelope spectrum with BPFO/BPFI reference lines drawn on top of the 3×/5× rotor harmonics is precisely the image that invites the confabulation this release refuses.

Nothing is gated and nothing is dropped. All records built by the producer are published. The envelope evidence still travels on every row as informational metadata (evidence_is_gate: false): computed_verdict, computed_snr, computed_dominant_line, evidence_tier, fault_freqs, line_resolvability. No record was filtered, relabelled or reordered by it, and nothing downstream should treat it as a filter — evidence_tier is present only for comparability with the gated siblings (MFPT / CWRU / XJTU). You can therefore re-derive every claim above from the published rows yourself.

Provenance & reproducibility

Generated deterministically by forge_agent/examples/mafaulda_bearing/convert.py (b3c9d23345) → forge_model/MAFAULDA_BEARING/convert_mafaulda_bearing.py (d6e0dab739); see provenance.json, which embeds the producer's full record under raw_form_provenance.

1,120 source sequences (5 s @ 50 kHz) → 2 fixed, non-overlapping 1 s windows per file at [1.0, 3.0] s — deterministic, evenly strided, away from both ends, and identical to the rotor sibling's sampling so the two families' windows are comparable record for record.

Each file's shaft rate is its filename (fr_source='filename' on every row). The tachometer channel is deliberately unused: FFT peak-picking on its pulse train returns pulse-train harmonics rather than the shaft rate, which would silently mis-place every fault line. Renderers are forge_tools defaults; spectrogram and scalogram stay full band (at 1 s / nperseg=256 the STFT bin is ~195 Hz at 50 kHz, so no zoom could resolve a fault-line ladder anyway) — all four representations are texture channels here, not readable spectra. 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, or as a documented negative control for label-free bearing evidence. The compute-then-check bearing tracks in this corpus are MFPT-*, CWRU-* and XJTU-*.
• Fault position is metadata, not part of the label. underhang vs overhang says where on the shaft the defective bearing sits; the four classes are the condition only. A position-aware task can be re-carved from the position field.
• One channel per record — the accelerometer at the fault position (underhang_radial for underhang and for healthy, overhang_radial for overhang; channel / channel_index on every row). The rig's other five axes exist upstream but showed no separation (see finding 3).
• Class balance follows the source: normal is by far the smallest class (49 source sequences vs ~1,071 seeded ones), and being a single assembly it is train-only.
• Speed is swept, not fixed (~12–62 Hz shaft rate, one speed per file). Speed is in fr_hz; a model can learn it from the image, so speed-stratified evaluation is advisable.
• The rig's rotor trees (imbalance / horizontal- and vertical-misalignment) are a different physics and a different conclusion — they are the sibling MAFAULDA-rotor release, which does ship a reasoning track.

Source & licence

Source: MAFAULDA — Machinery Fault Database, Signal, Multimedia and Telecommunications Lab (SMT), COPPE/Poli, Universidade Federal do Rio de Janeiro (http://www02.smt.ufrj.br/~offshore/mfs/); contact Felipe M. L. Ribeiro.

⚠️ Licence not stated upstream. As of 2026-07-28 the source page carries no licence, copyright notice, terms of use or citation requirement. This derived dataset is therefore distributed gated (manual approval) for research use, with attribution to SMT/COPPE/UFRJ. Clear the upstream licence with the maintainers before any onward redistribution or commercial use. If you are the rights holder and want this changed, please open a discussion on this repo.