Roles

Roles: canon repo — annot is the source label, kept machine-parseable as the gold for verification and reward parsing; there is no filled reasoning column and this repo is not itself a training view. Derived repos each state their own regime on their own card.

Paderborn KAt — bearing race damage from the envelope spectrum (reasoning track)

Part of the AI4Manufacturing FORGE corpus (Category C, task T-C1). Each record is the envelope spectrum of one 4 s vibration record from the Paderborn KAt bearing test rig, drawn with the outer-race (BPFO) and inner-race (BPFI) fault frequencies marked as reference lines. The verdict is a position: which marked line the energy sits on. That is what supports faithful compute-then-check chain-of-thought. reasoning is empty here; the PADERBORN-annotated sibling fills it.

Records: 1074 (splits {'test': 319, 'train': 755}); labels {'healthy': 469, 'outer_race': 328, 'inner_race': 277}; from 15 physical bearings; damage origin {'real': 781, 'artificial': 293}.

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 shipped in no repo (they are the whole of the raw form's excluded split): the corpus task is single-label, and a compound bearing has no single right answer. There is no ball and no cage class because the source documents none — the accelerated-lifetime runs produced race damage only. That is also why the figures mark four lines and not six, and why the query names BPFO and BPFI only: a marked BSF or FTF would point at a class the answer space does not contain.

⚠️ This repo holds 9 of the 23 damaged bearings, and that is the finding

Every record here passed the evidence gate, and its bearing passed an upstream one: a unit whose race line clears the screen on fewer than half of its ~80 records is routed to perception whole, because its few responding windows are windows that happened to light up rather than a spectrum that can be read. The bearings below are not defective files and their damage is documented; what they lack is readable evidence. The divide does not follow damage origin — artificial and real-pitted bearings appear on both sides — and no cause has been established.

• Routed out entirely (14): KA03, KA05, KA07, KA08, KA15, KA22, KA30, KI04, KI05, KI07, KI08, KI14, KI17, KI21. Their envelope spectra are published nowhere. They ship in PADERBORN-perception only, whose gate is per-record and whose four other representations still carry them.

metadata.unit_hit_rate and metadata.track_reason carry the upstream routing decision on every record, so you can re-cut on it.

Do not read this repo as a sample of the Paderborn dataset — it is the readable part of it, and a model's accuracy here says nothing about the routed-out units.

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 -annotated sibling)
cate / task	C / T-C1 (signal fault classification)
metadata	JSON string: representation, id, unit, condition, repeat, rpm, load_nm, radial_n, fs, fr_hz, channel, damage_origin, track, track_reason, unit_hit_rate, fault_freqs, computed_verdict, computed_snr, evidence_tier, baseline_used, baseline_drawn, line_evidence, image_sha256, split

Splits

train / test is a bearing-wise holdout: the test split is made of whole bearings, never seen in training under any operating condition. This is the split that matters here. Each bearing contributes up to 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.

After the evidence gate this track holds 1074 records over 15 bearings ({'test': 319, 'train': 755}); damage origin {'real': 781, 'artificial': 293}.

What the held-out set actually contains

class	bearings	which	records	share
outer_race	2	KA01 (artificial, 80), KA04 (real, 79)	159	50%
healthy	1	K001 (real, 80)	80	25%
inner_race	1	KI01 (artificial, 80)	80	25%

Answering the majority class for every record scores 50% on this test split (159/319). Report against that number, not against 33%.

Figures (style pool)

The spectra in this repo come from the figure-style pool: each record is drawn in one of 11 styles. What the style varies: colours, canvas, title placement and presence, tick spacing and minor ticks, grid, spines, fill, the fault-frequency marker drawn as a line or as a band, label placement and orientation, one colour per family plus a legend, font family, scientific notation, and tick direction. The style is picked by sha256("figstyle|" + image_path) % 11 -- deterministic, reproducible, and drawn independently of the opener / knowledge / skeleton pools.

What never varies is the content: the curve; the plotted band; which lines are marked (four on these figures -- the two race families plus 2xBPFO and 2xBPFI; the other four bearing sets in this corpus mark six, because they have families this one has no class for); and the y-limits, autoscaled once per record and reused by every style, so a peak is the same height whichever style drew it. No threshold bar and no noise-floor line is ever drawn, and no marker is styled differently for being the answer. The point is that a model cannot pass by memorising one house plotting style. The -perception repo's images are unchanged.

What pins these images. Three separate things produced them, and each carries its own evidence -- none of them stands in for another:

• the signal and its manifest: forge_agent commit not recorded;
• the curve that is plotted, and the producer parameters behind it (demodulation band, plotted fmax, which lines are marked): forge_agent 733a683. This is not the code that drew the PNGs;
• the published bytes: figure_manifest_sha256 = 203e42bb5cd0491fa8a61faf569d0891c3f44db62d5b3b07b2961571e791c2db -- the sha256 of the pool's per-image sha256 manifest (_final_sha256.txt, one line per image, checkable with sha256sum -c).

The styling layer itself (_figstyle_corpus/_src/style_pool.py) lives on the data volume and is not under version control, so no commit identifies it and none is offered: a reader who checked out the commit above would get the same curve drawn some other way. The manifest hash is what these images should be checked against. provenance.json records all three pins separately.

Provenance & the evidence gate

Generated deterministically by forge_agent/examples/paderborn/convert.py (commit not recorded) → forge_model/PADERBORN/convert_paderborn.py (754d363e51); full record in provenance.json. The raw form names its tool but no revision, so the signal pin above reads not recorded rather than a guess.

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.

The reasoning track then gates twice. Per record, it keeps confirmed only: a spectrum that shows no line cannot honestly carry a chain-of-thought that cites one. Per bearing, it keeps only units the upstream router sent here — a unit whose race line clears the screen on fewer than half of its ~80 records is routed to perception whole, because its few responding windows are windows that happened to light up rather than a spectrum that can be read, and a paragraph reasoning from one of those would teach a coincidence as evidence. metadata.track and metadata.track_reason carry that decision on every record. The perception configs apply neither gate at bearing level and keep confirmed + non-conflicting weak, so a damaged unit that stays silent is still represented somewhere.

Caveats

• Damaged bearings that never show a race line are not broken files. They respond on a minority of their 80 records — or on none — while others respond on more than 90 % under the same gate, the same four operating conditions and the same instrumentation. The header above names them for this build.
• The figure, the question and the paragraph use two families; the metadata records four. BPFO and BPFI are everything that is marked, everything the query names and everything the reasoning cites — this answer space has no rolling-element and no cage class. metadata.fault_freqs and metadata.line_evidence are the upstream measurement and still carry BSF and FTF: they are kept verbatim so every published number can be re-checked against the full record, rather than trimmed to match the picture. Metadata is never shown to the model at train time.
• False positives are left in and reported rather than tuned away. Healthy K002 at N15_M01_F10 (0.1 N·m load torque) reads as inner_race on part of its records. 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 dropped from the reasoning repos by the gate (a healthy image must stay quiet to qualify) and are present in PADERBORN-perception, which applies no such requirement.
• One source file is unreadable (N15_M01_F10_KA08_2.mat, Expecting matrix here) and is dropped, so KA08 contributes 79 raw records instead of 80.
• 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.
• 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 these repos.

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: 319 of 1,074 records (317 distinct images) are locked — by column: 0 by the cross-family manifest, 319 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; 319 locked by their own split: test. In words: 319 of the 1,074 records in this repository are evaluation material by their own metadata.split (test: 319) 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.