AI4Manufacturing/MAFAULDA-rotor

MaFaulDa — rotor fault classification from the shaft-order spectrum (reasoning track)

Part of the AI4Manufacturing FORGE corpus (Category C, task T-C1) and its first rotor dataset with a reasoning track. Each record is the raw shaft-order spectrum of a 1 s vibration window with the shaft harmonics marked — rotor faults are additive synchronous components, so unlike the bearing datasets this is not an envelope spectrum. reasoning is empty here; the MAFAULDA-rotor-annotated sibling fills it.

Records: 1258 (splits {'train': 1033, 'test': 225}); labels {'horizontal_misalignment': 241, 'imbalance': 491, 'normal': 80, 'vertical_misalignment': 446}; evidence_tier {'confirmed': 1258}.

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 / imbalance / horizontal_misalignment / vertical_misalignment
reasoning	chain-of-thought (empty here; filled in the -annotated sibling)
cate / task	C / T-C1 (signal fault classification)
metadata	JSON string: representation, features, harmonic_profile, harmonic_elevation, computed_verdict, family_indication, subtype_is_implanted_gold, evidence_tier, baseline_files, channel, fr_hz, severity, group_id, file, window_idx, fs, image_sha256, split

Splits

train / test, severity-wise (leakage-safe): one severity level is one physical assembly and lives wholly on one side; the highest severity of each fault class is held out. Windows never cross files.

What the evidence supports — three levels, all measured

This is the corpus's first rotor dataset with a reasoning track, and the honest scope of that evidence is narrower than the label set. Measured at build time on the overhang tangential channel against a speed-matched healthy baseline:

gold class	fires rotor_anomaly	1× elev (med)	2×	3×	family indication when fired
normal	0.184	0.98	1.02	0.95	misalignment 78% / unbalance 22% (n=18)
imbalance	0.737	2.76	2.50	1.01	misalignment 57% / unbalance 43% (n=491)
horizontal_misalignment	0.612	1.27	1.79	1.69	misalignment 87% / unbalance 13% (n=241)
vertical_misalignment	0.741	1.13	2.85	1.18	misalignment 91% / unbalance 9% (n=446)

1. Binary anomaly (gated). "Is this window anomalous versus the same rig's own healthy baseline?" This is what evidence_tier gates on; the reasoning repo keeps confirmed only.
2. Family indication (not a gate, and directional). A fired window that looks misalignment-like. Misalignment files read misalignment ~87–91% of the time — but imbalance files split ~57/43, because added mass lifts 2× on this rig too. It is not an imbalance detector, and a chain-of-thought written over these records must not claim it is.
3. Subtype is implanted gold. Horizontal vs vertical misalignment is not separable by this evidence; those records carry subtype_is_implanted_gold: true. The distinction comes from the rig operator's documented setup, not from the signal.

Healthy false-positive rate ≈18% — structural, not a defect: MaFaulDa sweeps 12–61 Hz and every file runs at a different speed, so a healthy window is scored against neighbours ~1 Hz away and speed-tracking resonances read as elevation. Those windows tier weak and drop from the reasoning track. The threshold was deliberately not tuned against the gold labels.

Provenance & reproducibility

Generated deterministically by forge_agent/examples/mafaulda_rotor/convert.py (b3c9d23345) → forge_model/MAFAULDA_ROTOR/convert_mafaulda_rotor.py (d6e0dab739); see provenance.json.

Each file's shaft rate is its filename (re-verified at build: median healthy raw 1× SNR 17.4); the tachometer channel is deliberately unused because FFT peak-picking on it returns different pulse-train harmonics. Baselines are speed-matched: the three nearest-speed healthy files, each profiled at its own shaft rate then combined by median — profiling them at the target's rate under-measures the baseline ~30% and inflates every elevation. The three files actually used are recorded on every row (baseline_files), so the protocol is auditable per record.

Caveats

• Bearing faults are excluded. The same rig also ships 1,071 seeded bearing sequences (underhang/overhang). They are a different physics (impact modulation, not shaft-synchronous) and are handled separately — on this rig their characteristic orders (BPFO ≈ 3×, BPFI ≈ 5×) collide with the rotor's own harmonics, so they cannot support this kind of evidence.
• One rig, one channel. All records are the overhang tangential accelerometer. The axial channel showed no separation here and the radial channel none at all — measure before assuming a channel carries misalignment.
• Class balance follows the source (imbalance and vertical misalignment dominate; normal is the smallest class and, being the evidence baseline, sits wholly in train).
• Split is severity-wise: one severity level = one physical assembly. test holds out the highest severity of each fault class, so it measures generalization to an unseen severity.

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/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.