First gear dataset in the AI4Manufacturing FORGE corpus (Category C, task T-C1). Each record is the envelope-modulation spectrum of a 4 s planetary-gearbox vibration window with the gear-train reference orders marked (shaft 1x/2x/3x, planet carrier, sun-gear defect) — the representation for faithful compute-then-check CoT. What the picture is for: on a faulted gear the sun-fault family (5/6, 1/6, 1/4 × shaft) overtakes the integer shaft orders; on a sound one the integer orders win. The question is binary: healthy or gear_fault. It used to be a five-way subtype question; Why two classes and not five below has the measurements that changed it, and the subtype is still in metadata.fault_subtype. reasoning is empty; the SEUG-annotated sibling fills it.

Records: 382 (splits {'train': 305, 'test': 77}); labels {'gear_fault': 311, 'healthy': 71}; evidence_tier {'confirmed': 382}. The implanted subtype behind those labels (metadata.fault_subtype, not the answer): {'health': 71, 'chipped': 80, 'miss': 80, 'root': 78, 'surface': 73}.

Why two classes and not five

This repo used to ask for the implanted subtype (health / chipped / miss / root / surface). It asks for presence or absence instead, for three reasons.

1. Measured: the four fault subtypes do not survive a change of operating condition. The rig was run at two speed-load cells (20 Hz-0 V and 30 Hz-2 V). Train a simple classifier on one cell, test on the other (leave-one-condition-out), five-way, chance 0.200:

what the classifier was given	hold out 20 Hz-0 V	hold out 30 Hz-2 V
the 6 low-order modulation observables already in this repo's metadata	0.44	0.33
mesh-frequency (GMF) sidebands, 20 features	0.19	0.28
all 8 channels x 144 dimensionless ratios	0.29	0.21
angle-domain TSA + residual shape, 60 features	0.22	0.24

Four representations resting on different physics -- the low-order modulation observables this repo already ships, mesh-frequency sidebands, dimensionless ratios across all eight channels, and angle-domain synchronous averaging with residual-shape features -- and not one of them clears chance on both held-out cells. Going channel by channel does not rescue it either (best, Channel5, averages 0.33). Published five-way accuracies on this dataset are near-perfect because they are measured within a recording, where a classifier can key on that recording's own operating point.

2. The literature says these are not five independent types. The reference taxonomy for localised gear damage has two mechanisms, cracks and spalls, and the five SEU names do not partition onto them. Verbatim: "the case of a missing tooth ... can be considered as the ultimate result of a developing crack, as could many cases of 'chipped teeth', where the crack does not grow across the whole tooth. On the other hand, chipping of teeth can occur as a result of a sudden local trauma (foreign object damage) and is likely to cause symptoms similar to those of spalls" (D. Peng, W. A. Smith, P. Borghesani, R. B. Randall, Z. Peng, Mechanical Systems and Signal Processing 127:531-550, 2019, doi:10.1016/j.ymssp.2019.03.024, §1). So root and miss are two severities of one crack, surface is the spall branch, and chipped straddles both -- the same name covers a crack that stopped short and an impact that behaves like a spall. Asking a model to separate root from miss is a severity regression wearing a classifier's clothes, and asking it to place chipped is asking for a distinction the word itself does not make.

3. The method that would NAME a tooth fault cannot be run on this data. Naming a gear-tooth fault from vibration is an angle-domain job: time-synchronous-average the signal against the suspect gear, then read the shape of what is left over. On a planetary stage that does not work. From the US Army Research Laboratory and NASA Glenn, on the planetary stage of an OH-58C main transmission: "The dynamics of the planet gear and sun gear are commensurate; thus, time-synchronous averaging could not be used as a separation tool as it may be for the pinion and bevel gear components" (A. Hood, K. LaBerge, D. Lewicki, D. Pines, Vibration Based Sun Gear Damage Detection, ASME DETC2013-12361, PTG 2013; NASA NTRS 20140005560). The same paper's method section says what it takes to do it at all: a once-per-revolution tachometer pulse on the carrier/output shaft to cut the signal into carrier cycles, and the ring tooth count to set the angular interpolation length (plus the hunting-tooth ratio, itself a function of the tooth counts, to decide which cycles may legitimately be averaged together). This release ships no tachometer channel, and this rig's tooth counts are not published -- see the data caveats and the reference-line audit.

What the evidence does support is the binary question. The label-free detector in the pipeline (mesh_modulation) has always answered fault_evident / healthy, and it does so at both operating conditions. Its agreement with the gold is the evidence_tier histogram quoted above.

The perception track was measured separately

The four perception renders are a different representation from the modulation spectrum, so the 5-way question was put to them, not inherited. A simple classifier sees exactly what a VLM sees -- the PNG, greyscaled to 64x64, no metadata and no signal -- PCA to 40 components, multinomial logistic regression, leave one operating condition out, 5-way, chance 0.200:

perception render	hold out 20 Hz-0 V	hold out 30 Hz-2 V
`spectrogram`	0.670	0.209
`scalogram`	0.415	0.204
`waveform`	0.405	0.325
`reshaped`	0.180	0.189

reshaped is at chance. The other three are not, and one cell is striking: at 32x32 the spectrogram trained on the 30 Hz cell scores 1.000 on the 20 Hz cell. That number is reported here rather than buried, and it is still not a five-way task -- three things about it:

It runs one way only. The same model trained on the 20 Hz cell scores 0.209 on the 30 Hz cell, which is that test set's majority class: it puts every window in one bin. A rule that works in one direction and collapses in the other is not describing the fault.
It is five recordings. There is one continuous recording per (class, condition), so a held-out condition is 5 independent items, not 200 windows. Counted that way the spectrogram gets 5/5 recordings right in the good direction and 1/5 in the other (3/5 and 1/5 at 64x64). There is no error bar to put on that.
It does not survive nuisance choices. Over 64 runs (4 renders x 2 image sizes x 2 PCA widths x 4 classifiers) the mean is 0.313 and the spread is 0.169-0.703; inside the spectrogram alone a single hold-out ranges 0.209-1.000.

It is not a trivial loudness cue either -- the window's RMS alone scores 0.200 / 0.162 and the image's mean and standard deviation score 0.275 / 0.257 -- which is the point: what transfers is the recording's own texture, and it transfers to one neighbour and not the other. So the perception configs carry the same two labels as the reasoning track.

For contrast, the binary question on the same windows, answered by the label-free detector with no training and no labels at all: 0.970 at 20 Hz-0 V and 0.940 at 30 Hz-2 V (0.9550 overall; healthy recall 1.000 and 0.775, fault recall 0.963 and 0.981). That is the asymmetry the label change is built on.

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 gear condition, two classes: `healthy` / `gear_fault`
`reasoning`	chain-of-thought (empty here; filled in the `-annotated` sibling)
`cate` / `task`	`C` / `T-C1` (signal fault classification)
`metadata`	JSON string: `fault_subtype` (the implanted subtype -- read the warning below before using it), representation, condition, file, window_idx, start_sample, channel, fs, fr_nominal, fr_used, fr_source, planetary, gear_lines, computed_verdict, computed_score, integer_score, family_obs, evidence_tier, image_sha256, split

Reference lines: what they are called, and what was re-checked

The markers are labelled on the image in words, because the image is read by a model with no glossary. metadata.gear_lines keeps the trade's abbreviations as its keys — that is a data field, and renaming it would break downstream code. This is the bridge:

label drawn on the image	key in `metadata.gear_lines`	frequency	what it means
`planet carrier`	`fc1`	1/6× shaft	how fast the planet carrier turns
`sun-gear defect`	`sunfault1`	5/6× shaft	how often a given spot on the sun gear is struck by a planet -- the sun's rotation rate relative to the carrier
`shaft 1x`	—	1× shaft	the input shaft's own rotation rate
`shaft 2x`	—	2× shaft	twice the shaft rate
`shaft 3x`	—	3× shaft	three times the shaft rate
(not drawn — off the axis)	`GMF1`	16.6667× shaft	stage-1 gear mesh: teeth going past per second
(not drawn — off the axis)	`GMF2`	3.1111× shaft	stage-2 gear mesh (unverified -- see below)

Earlier renders of this dataset drew fc1, sunfault1 and 1x/2x/3x instead. Same lines, same frequencies, same underlying data — only the marker text changed.

The dashed markers on the reasoning image are predictions, placed from the measured shaft rate and the derived gear train. They are not observations, and a reader of the image is meant to check them. Re-measured on this dataset's own spectra (Welch over the whole recording, 2^18-point Hann blocks, ~0.02 Hz bins; prominence = the peak in a ±0.4 % order window over the median of the surrounding ±5 %):

The mesh order is real. Stage-1 GMF is predicted at 16.6667× shaft and is observed at 16.6739× (20 Hz cell) and 16.6636× (30 Hz cell) -- within 0.05 % at both speeds, prominence ×7 and ×17. It is nonetheless not a tall peak: it ranks 46th and 21st among the peaks below 2 kHz, at 8.6 % and 15.5 % of the tallest one. The tallest peaks are integer shaft harmonics (139.7 Hz = 7x shaft in the 20 Hz cell, 269.6 Hz = 9x shaft in the 30 Hz cell) -- different orders, nearly the same hertz, i.e. a structure resonance in the 140-280 Hz region, not a gear line. Anyone comparing gear_lines.GMF1 against the biggest thing in the spectrum will conclude it is wrong; it is not.
The planet carrier marker (gear_lines.fc1) is the weak one, and it fails at the faster speed. The derivation's stated evidence for stage 1 included "a carrier line at exactly fr/6" — the line this marker points at. In the raw spectrum there is no such line at either speed: prominence ×0.7-1.5 against its own neighbourhood, which is the floor. In the envelope-modulation spectrum the image actually plots, score the marker the way the detector does and compare it with 200 decoy frequencies drawn from the same band: it lands at percentile 63-100 in the 20 Hz cell but only 27-78 in the 30 Hz cell -- at 30 Hz it is an ordinary frequency, below the median decoy in three of the five classes. The sun-gear defect marker (gear_lines.sunfault1) at 5/6× shaft, by contrast, is the strongest thing in its band in all ten class × condition cells (100th percentile everywhere; raw-spectrum prominence ×12.9-19.5) -- including in the healthy windows, because it is a kinematic line of this gear train and not a fault line. What separates the two classes is the contrast: on a faulted gear the sun-fault family (5/6, 1/6, 1/4 × shaft) overtakes the integer shaft orders, and on a sound one the integer orders win. In the 30 Hz cell the sun-fault line is often the tallest peak in a healthy window too, which is where the detector's healthy recall goes (0.775 there, against 1.000 at 20 Hz). Both markers are drawn on every image here, healthy records included — the same convention as the BPFO/BPFI markers on the bearing sets: the line says where such a defect would show, it does not claim one is there. Read planet carrier as a hypothesis this data half-refutes rather than as a located line -- and note that fr/6 is also one of the three orders the detector sums, so at 30 Hz that term is contributing noise.
The hertz values in gear_lines carry a small speed bias. fr_used is refined from each 4 s window, whose spectrum has 0.25 Hz bins, so it lands at 0.26 % (20 Hz cell) and 0.11 % (30 Hz cell) above the shaft rate measured over the whole recording (19.9414 Hz and 29.9609 Hz). Every line scales with it: GMF1 is reported as 333.25 Hz in the 20 Hz cell where the whole-record measurement puts it at 332.4 Hz. Harmless for the plotted low-order markers (0.04 Hz, against a 0.4 Hz matching tolerance); worth knowing before quoting gear_lines to four digits.
The tooth counts are a derivation, and only half of it is corroborated. Nobody publishes this rig's tooth counts: the manufacturer's spec page gives the gearing as "2-stage, 27:1 ... 4 planet stage1 and 3 planet stage 2" and stops there; the dataset paper and the release repo give none; and of 17 open-access papers using this dataset, none states a tooth count, a reduction ratio, or even which gear the faults were implanted on. What does exist is a table published by a different lab running the same DDS model (C. Liu, G. Cheng, X. Chen, Y. Pang, Sensors 18(5):1523, 2018, doi:10.3390/s18051523). Stage 1 agrees exactly with the set derived here -- Zs 20 / Zp 40 / Zr 100 -- which is an independent confirmation of the only gear order this repo's images depend on. Stage 2 does not: published 28/36/100 against 24/30/84 here. This dataset cannot settle it -- the derived stage-2 mesh order (3.111×) shows prominence ×4.2-6.5 at both speeds, the published one (3.646×) shows ×6.5-8.3 at 30 Hz but only ×1.7-2.1 at 20 Hz. Treat gear_lines.GMF2 as unverified. No image in this repo draws it, and no label depends on it.

None of this changes a label: the labels are the implanted-fault filenames, and the binary task above does not depend on the gear train being right.

Provenance & reproducibility

Generated deterministically by forge_agent/examples/seu/convert.py (0574e0e25d) → forge_model/SEUG/convert_seug.py (9448e33334); see provenance.json.

Gold = filenames (the files' internal Title fields are provably stale operator templates); the gear conditions are physically implanted and steady-state, so every window carries its file's condition. The gear-train constants (2-stage planetary 20/40×4/100 → 24/30×3/84, 27:1) were derived from this dataset's own spectra, not read off a spec sheet: the manufacturer sells the rig as reconfigurable and publishes no tooth counts. Which parts of that derivation hold up against the data — and which one does not — are set out under Reference lines: what was re-checked. Full chain and confidence grades in provenance.json (planetary_derivation).

Caveats

The four implanted subtypes are still shipped, in metadata.fault_subtype — but do not build a five-way task on them without reading Why two classes and not five first. Within a single recording they separate easily and the accuracy will look excellent; across the two operating conditions four different representations all sit at chance. The field is kept because a severity study, or a deliberately within-condition study, may legitimately want it. A cross-condition five-way number computed from this field is not a generalization result.
The two classes are not balanced. The rig recorded one healthy run against four faulted runs per operating condition, so answering gear_fault every time already scores 0.814. Report balanced accuracy or per-class recall; plain accuracy on this split is close to meaningless.
The reasoning track is gated on a label-free detector agreeing with the gold, and the task is now in that same binary label space. So the SEUG track is, by construction, the subset of windows on which mesh_modulation already gets the answer right — that is the point of a compute-then-check track (the evidence is in the picture), but it means a score there is not an estimate of performance on unscreened gearbox data. SEUG-perception keeps the records the detector missed (see the conflict rule) and is the harder set.
Conflict rule (binary): a weak record is dropped only when the detector claims a fault on a healthy record; a quiet detector on a gear_fault record is benign non-detection and is kept in perception.
Split is time-stratified per file (first 80 % of each recording → train, last 20 % → test). It is a split within one recording, not between specimens — see the next block.
Two operating conditions (20 Hz-0 V and 30 Hz-2 V motor speed-load) are both included, tagged in metadata.condition. Holding one out is the honest generalization test on this dataset, and it is much harder than the shipped split.

What the public release does not contain (re-verified 2026-09-20)

The files are truncated to a quarter of the recording their own header describes. Every gearset CSV declares Total Data Rows 4194304 (= 1024 blocks × 4096 samples, the DAQ's Number of Blocks × Spectral Lines × 2.56) and then carries 1,048,560 data rows — 25.0 %, about 205 s at 5120 Hz. The missing three quarters are simply not in the public release. What is there is continuous: across the 255 interior 4096-sample block boundaries the mean sample-to-sample step is 0.98-1.06× the mean step elsewhere and the largest boundary step stays below the 99.99th percentile of ordinary steps, so it can be treated as one unbroken record.
There is no keyphasor or tachometer channel. All eight channels are taken: 1 motor vibration, 2-4 planetary gearbox x/y/z, 5 motor torque, 6-8 parallel gearbox x/y/z — seven accelerometers and one torque channel. Shaft rate has to be estimated from the vibration itself (fr_source, fr_used), and the once-per-revolution carrier pulse that angle-domain planetary diagnostics needs does not exist here.
One continuous recording per (class, condition) cell — ten files for five conditions × two operating points. Every window, in every split, comes from inside one of those ten recordings. There is no second assembly, no re-fit, no repeated run: nothing in this dataset can test whether a model survives the gearbox being taken apart and put back together, which is the failure mode that matters in the field.

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: DAQ header × 2.56 convention + shaft combs at nominal in both conditions]. The release's dataset/ folder (CWRU fan-end copies) is excluded — CWRU is published separately from its original source.

Downloads last month: 59

Total file size:

28.9 MB