AI4Manufacturing/SEUB-perception

SEU bearingset — perception representations (visual grounding)

The SEU bearingset windows rendered as perception images — one HF config per representation, for the foundation model's visual grounding (subtype discrimination here is texture-learnable, not physics-nameable — see caveats).

Configs

load_dataset("AI4Manufacturing/SEUB-perception", "spectrogram")

config	records	splits
spectrogram	398	{'train': 318, 'test': 80}
scalogram	398	{'train': 318, 'test': 80}
waveform	398	{'train': 318, 'test': 80}
reshaped	398	{'train': 318, 'test': 80}

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: health / inner_race / outer_race / ball / inner_outer_comb
reasoning	chain-of-thought (empty here; a -annotated sibling may fill it later)
cate / task	C / T-C1 (signal fault classification)
metadata	JSON string: representation, condition, file, window_idx, start_sample, channel, fs, fr_nominal, fr_used, fr_source, planetary, gear_lines, computed_verdict, computed_score, family_order, family_hz, evidence_tier, image_sha256, split

Provenance & reproducibility

Generated deterministically by forge_agent/examples/seu/convert.py (a990b2ef69) → forge_model/SEUB/convert_seub.py (8892ffb2db); see provenance.json.

Gold = filenames (the files' internal Title fields are provably stale — ball_30_2 is titled "outer_30_2" yet its data is demonstrably distinct); the five bearing conditions are physically implanted in the DDS parallel gearbox (kit menu matches the class set; comb = documented inner+outer combination) and steady-state. The bearing model/geometry is unpublished, so the evidence detector names no BPFO/BPFI: it reports the strongest modulation family not attributable to the known gear train (whose constants were derived from this dataset's own spectra; stage-1 grades high, stage-2 moderate — full chain in provenance.json planetary_derivation).

Caveats

• No named bearing physics. The parallel-gearbox bearing model is unpublished; the label-independent detector (modulation_family) attests that abnormal modulation is present at orders not attributable to the excluded STAGE-1 gear-train set (carrier half-harmonics, planet spin, shaft integers). Stage-2 planetary orders are deliberately NOT excluded (excluding them would blind genuine sub-carrier bearing modulation) — so a confirmed family can in principle sit on an unmodelled stage-2 line; the tier is a binary amplitude-anomaly attestation, not a named-fault identification. Its operating point favours a clean healthy class (measured health false-alarm 0% at prominence 20); the cost falls on the FAULT side — at this threshold no fault class reaches confirmed (see provenance detector_operating_point for the measured per-class fractions), which is exactly why this release is perception-only. Subtype discrimination is learnable from these signals (deep-learning literature) but not physics-nameable.
• Conflict rule (binary): weak records are dropped only when the detector claims a fault on a health record (none occurred at this operating point); a quiet detector on a fault record is benign non-detection (kept in perception).
• Split is time-stratified per file (first 80% → train, last 20% → test): ONE physical specimen per (condition, speed-load) cell — no unit-wise split exists; cross-specimen generalization cannot be evaluated from this dataset.
• Two operating conditions (20 Hz-0 V, 30 Hz-2 V) included with condition metadata. Two absent-tier windows (no modulation energy at all) are dropped, hence 398 records per config rather than 400.

Source & license

Source: SEU gearbox dataset — Southeast University, Drivetrain Dynamics Simulator (SpectraQuest/Sumyoung DDS). Authors' research release: github.com/cathysiyu/Mechanical-datasets (no LICENSE file — cite the paper): S. Shao, S. McAleer, R. Yan, P. Baldi, IEEE Trans. Industrial Informatics 15(4):2446–2455, 2019 (DOI 10.1109/TII.2018.2864759). fs = 5120 Hz [evidenced]. The release's dataset/ folder (CWRU fan-end copies) is excluded.