AI4Manufacturing/PADERBORN-perception

Roles

Roles: perception view of PADERBORN — annot is the source label (healthy / inner_race / 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) — 2,319 records each, and the query states which encoding the record shows. The reasoning column is empty on all 9,276 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.

Paderborn KAt — bearing race damage, perception representations (grounding track)

Part of the AI4Manufacturing FORGE corpus (Category C, task T-C1). Four image renderings of the same 4 s vibration record from the Paderborn KAt bearing test rig, one HF config each.

Records: 9276 across 4 configs (2319 records each, from all 29 shipped bearings); labels {'healthy': 480, 'outer_race': 959, 'inner_race': 880}; splits {'test': 400, 'train': 1919}.

config	image
spectrogram	short-time Fourier transform (time × frequency)
scalogram	Morlet continuous-wavelet transform (time × scale)
waveform	the raw time trace
reshaped	samples arranged into a 2-D grayscale grid

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

The reasoning counterpart — the envelope spectrum, the representation the race frequencies are read off — lives in AI4Manufacturing/PADERBORN. It holds only the 1074 records whose fault line is actually visible.

⚠️ No evidence gate is applied here, deliberately

Every non-excluded record ships, including the 14 damaged bearings whose fault line never appears. That is not an oversight. These four views make no claim about where the energy sits — they carry texture, not a cited line — and the labels are the source's own documented damage. A bearing whose race line is unreadable is still a damaged bearing.

So metadata.evidence_tier is informational on this track: nothing was dropped, relabelled or reordered by it. Use metadata.track (reasoning / perception) and metadata.unit_hit_rate if you want to compare the readable and unreadable halves — that comparison is the most interesting thing in this dataset and it is only possible because both halves ship.

What this track will not give you: the fault frequency. If you want to know whether a model can name the damaged race from evidence rather than from texture, use the reasoning repo.

Rig & operating conditions

Chair of Design and Drive Technology, Paderborn University. An electric motor drives a test bearing module through a torque-measuring shaft against a load motor; the vibration is read by a piezoelectric accelerometer at the top end of the bearing housing and sampled at 64 kHz. Thirty-two 6203 bearings are measured: 6 healthy, 12 with artificially introduced damage (EDM, drilling, electric engraving) and 14 damaged by accelerated lifetime testing — real pitting and plastic deformation, not a machined notch.

Every bearing is run at four operating conditions, each a combination of three knobs, and each condition is repeated 20 times for 4 s:

condition code	speed	load torque	radial force
N15_M07_F10	1500 rpm	0.7 N·m	1000 N
N09_M07_F10	900 rpm	0.7 N·m	1000 N
N15_M01_F10	1500 rpm	0.1 N·m	1000 N
N15_M07_F04	1500 rpm	0.7 N·m	400 N

Answer space

healthy / inner_race / outer_race. The three KB bearings carry combined inner- and outer-race damage and are not shipped in either repo (240 records): the corpus task is single-label, and a compound bearing has no single right answer. There is no ball or cage class because the source documents none — the accelerated-lifetime runs produced race damage only.

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
reasoning	chain-of-thought (empty here; filled in the PADERBORN-annotated sibling)
cate / task	C / T-C1 (signal fault classification)
metadata	JSON string: unit, condition, repeat, rpm, load_nm, radial_n, fs, fr_hz, channel, damage_origin, representation, track, track_reason, unit_hit_rate, computed_verdict, computed_snr, evidence_tier, evidence_measured_on, baseline_used, baseline_drawn, image_sha256, split

Splits

train / test is a bearing-wise holdout: the test split is K001, KA01, KA04, KI01, KI04 — whole bearings, never seen in training under any operating condition. This is the split that matters here. Each bearing contributes 80 records (4 conditions × 20 repeats) that are near-repeats of one another, so a random record split reports how well a model recognises this unit, not this damage.

What the held-out set actually contains

class	bearings	which	records	share
healthy	1	K001 (real, 80)	80	20%
inner_race	2	KI01 (artificial, 80), KI04 (real, 80)	160	40%
outer_race	2	KA01 (artificial, 80), KA04 (real, 80)	160	40%

Answering the majority class for every record scores 40% on this test split (160/400). Report against that number, not against 33%.

This repo holds every shipped bearing: 2319 records × 4 representations.

Provenance & the evidence gate

Generated deterministically by forge_agent/examples/paderborn/convert.py (0574e0e25d) → forge_model/PADERBORN/convert_paderborn.py (9448e33334); full record in provenance.json.

Labels come from the source's own bearing catalogue. An evidence gate — a label-independent computation on the raw vibration — decides what may enter the reasoning track: a spectral-kurtosis kurtogram picks the demodulation band, a Hilbert envelope is taken, and the envelope spectrum is searched for the BPFO and BPFI line families. The search never sees the label.

The kurtogram is capped at 12 kHz. Left unbounded on a 64 kHz record it selects a band above 12 kHz on nearly every file: that region is the sensor's own noise floor, which is impulsive and therefore scores high kurtosis while carrying no bearing signal. The cap is a property of this rig's instrumentation, set once and applied to every record, healthy and damaged alike.

Each line is scored against a per-condition healthy baseline built from the training bearings only, so a line counts as present when it stands above what the same operating point looks like on a good bearing — not above an absolute number carried over from another rig.

On this track the gate is none (labels are the source's documented damage; evidence rides along as metadata).

Caveats

• 14 damaged bearings never show a race line, and they are not broken files. KA03, KA05, KA07, KA08, KA15, KA22, KA30, KI04, KI05, KI07, KI08, KI14, KI17, KI21 respond on fewer than half of their 80 records; six others respond on more than 90 % under the same gate, the same four operating conditions and the same instrumentation. The split does not follow damage origin — artificial and real-pitted bearings appear on both sides — and no cause has been established. These units ship perception-only; their envelope spectra are not published at all.
• 11 false positives, all on one bearing at one operating point. Healthy K002 at N15_M01_F10 (0.1 N·m load torque) reads as inner_race on 11 of its 20 records. This is left in and reported rather than tuned away. It is not explained by wear: the source's own datasheet gives that bearing 19 h of run-in against over 50 h for K001, which produces no false positive at all. Those records are excluded from the reasoning repo by the evidence gate (a healthy image must stay quiet to qualify) and are present in the perception repo, which applies no gate.
• One source file is unreadable (N15_M01_F10_KA08_2.mat, Expecting matrix here) and is dropped, so KA08 contributes 79 records instead of 80. Logged in provenance.json.
• Artificial and real damage are both present and are not interchangeable. metadata.damage_origin is artificial (EDM / drill / engraver) or real (accelerated lifetime test). A machined notch is a cleaner impulse source than natural pitting; training on one and reporting on the other is a different experiment, and the field is provided so you can run it deliberately.
• The evidence gate is measured on the SIGNAL, not on the image. The line search reads the raw vibration; it never opens the PNG. So a confirmed tier says this signal supports the label, not this image shows it. Every record carries evidence_measured_on: "signal". Image-side readability has not been measured.
• metadata.condition is a real confounder, not a nuisance field. The four conditions differ in speed, torque and radial force, and all four are present for every bearing. If you split at random, a model can score by recognising the operating point. The shipped split holds out whole bearings; keep it, or make your own the same way.
• Only one accelerometer channel is rendered (vibration_1). The source also ships motor phase currents at 64 kHz; those are a separate release and are not in this repo.

Source & license

Source: KAt-DataCenter, Chair of Design and Drive Technology, Paderborn University (mb.uni-paderborn.de/kat/datacenter). Please cite: Lessmeier, Kimotho, Zimmer & Sextro, Condition Monitoring of Bearing Damage in Electromechanical Drive Systems by Using Motor Current Signals of Electric Motors: A Benchmark Data Set for Data-Driven Classification, European Conference of the PHM Society 2016.

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: 1,600 of 9,276 records (1,584 distinct images) are locked — by column: 0 by the cross-family manifest, 1,600 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; 1,600 locked by their own split: test. In words: 1,600 of the 9,276 records in this repository are evaluation material by their own metadata.split (test: 1,600) 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.