AI4Manufacturing/CWRU-perception

Roles

Roles: perception view of CWRU — annot is the source label (ball / inner_race / normal / outer_race), kept machine-parseable as the gold for verification and reward parsing; the model reads query + image, where the repo ships a bearing's vibration in four image encodings as four equal-sized configs — reshaped (consecutive samples arranged as the rows of a grayscale square), scalogram (a continuous-wavelet time-scale view), spectrogram (a short-time Fourier transform) and waveform (the time-domain trace) — 762 records each, and the query states which encoding the record shows. The reasoning column is empty on all 3,048 records, so this repo is not itself a training view for chain-of-thought and trains annot directly. metadata is provenance (split keys, source identities) and must never be fed to a model.

CWRU — perception representations (visual grounding)

The same CWRU bearing windows rendered as perception images — one HF config per representation. A frozen VLM vision encoder carries real but modest class signal from all of these (shuffle-controlled probe): on a within-condition split it scores high, but under a strict bearing-wise split (unseen bearings) it is only slightly above chance — generalizing CWRU fault type across bearings is hard for every method, including a purpose-built 1D-CNN. So these are included as representation-diversity / grounding data for the foundation model (not as a high-accuracy CWRU classifier). Unlike the CWRU (spectrum) repo, they are not for compute-then-check CoT (reasoning stays empty) — the discriminative signal is non-verbal texture.

Configs

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

config	records	splits
spectrogram	762	{'train': 629, 'test': 133}
scalogram	762	{'train': 629, 'test': 133}
waveform	762	{'train': 629, 'test': 133}
reshaped	762	{'train': 629, 'test': 133}

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 fault class: normal / inner_race / outer_race / ball
reasoning	chain-of-thought (empty here; a planned -annotated sibling will fill it — not yet published)
cate / task	C / T-C1 (signal fault classification)
metadata	JSON string: representation, features, fault_freqs, computed_verdict, computed_snr, evidence_tier, sr_nondiagnosable, anomaly, size_in, load_hp, or_position, bearing, bearing_group, channel, rpm, fs, fr_hz, file_number, window_idx, image_sha256, split

Provenance & reproducibility

Generated deterministically by forge_agent/examples/cwru/convert.py (0574e0e25d) → forge_model/CWRU/convert_cwru.py (9448e33334); see provenance.json for the full record.

Cleaning (all encoded in the pinned code): excluded the NTN 0.028″ files 3001–3008 (wrong bearing geometry) and the corrupt 191; flagged Smith & Randall (2015) non-diagnosable (sr_nondiagnosable) and acquisition-anomaly files; bearing-wise leakage-safe split (all loads + both sample rates of a physical bearing stay on one side). Evidence-gated, conflict-free release: a computed evidence_tier (from the label-independent envelope-spectrum detector, cross-checked against the Smith & Randall expert grades) curates the published set so every image supports — and never fights — its label. The spectrum/reasoning repo keeps confirmed only (faithful compute-then-check CoT); the perception repo keeps confirmed + non-conflicting weak (weak records where the detector confidently found a different pattern than the gold are dropped).

Caveats

• Ball is the hard class — Smith & Randall (2015) grade most CWRU ball faults non-diagnosable (intermittent load-zone contact + smeared BSF), so many ball windows are weak/absent. The evidence gate keeps only the diagnosable ones, so the reasoning/spectrum track has the fewest ball records (confirmed-only) while perception also admits weak ball. A property of the data, not the converter.
• Class balance — after gating, normal is the minority (CWRU ships few baseline files) and outer_race the majority; per-class counts are in the header above.
• Eval: use the provided bearing-wise split; the within-condition (per-load) split inflates accuracy.

Source & license

Source: Case Western Reserve University Bearing Data Center (engineering.case.edu/bearingdatacenter), openly available for research. Diagnosability grades and anomaly flags from W. A. Smith & R. B. Randall, MSSP 64–65 (2015) 100–131; leakage-safe split per J. Hendriks et al., MSSP 169 (2022) 108732.

Overlap / de-duplication (§8)

Cross-family evaluation lock — metadata.eval_lock (stamped 2026-09-20; manifest revision fe6e286912b0, generated 2026-09-08). Every record of this repository, locked or not, carries metadata.eval_lock, computed by forge_model/common/overlap.py::Overlap.stamp_for against common/overlap_manifest.json at that revision — so within this repository the absence of the key cannot occur. Shape: {"locked": bool, "against": [{"repo": …, "split": …}, …], "own_split": …, "manifest_revision": …, "manifest_generated": …}. locked is true when the image is evaluation material anywhere in the corpus; against names every repository and split in which it is (sorted; [] when not locked; it includes the record's own family where that is so); own_split marks a record locked by its own split. The per-record field is the authority — the count here is quoted once, at this revision, and a later manifest may change it: 532 of 3,048 records (532 distinct images) are locked — by column: 0 by the cross-family manifest, 532 by their own split, 0 both ways and counted once; counterparts (records per counterpart; a record can appear under several): none — every lock here is by the record's own split; 532 locked by their own split: test. In words: 532 of the 3,048 records in this repository are evaluation material by their own metadata.split (test: 532) and sit inside the HF split named test / train — under the uniform-split convention the HF split name is a container name, and metadata.split together with metadata.eval_lock carries the truth; a train pool must exclude them. A stamp whose manifest_revision differs from the current manifest is stale, not wrong — recompute it (Overlap.stamp_is_current); a record with no stamp has not been checked against the corpus as it now is. Overlap.partition / assert_train_pool_clean read the field: a train pool built from this repository must exclude every locked record.