Records: 1600 across 4 configs (400 runs each); labels {'1_broken_bar': 80, '2_broken_bars': 80, '3_broken_bars': 80, '4_broken_bars': 80, 'healthy': 80}.

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/IEEE-BRB-perception", "spectrogram")

What these four views can and cannot answer

The broken-bar signature is a pair of lines 2·s·f_line either side of the supply frequency — 0.71 Hz apart at the lightest load, 5.66 Hz at full — sitting 30–50 dB below it. The render settings decide whether that survives onto the image, and the toolkit defaults did not: they were tuned for broadband bearing vibration, where time resolution matters and the band of interest is the top octaves of fs. Carried onto 50 kHz stator current they gave a 195 Hz spectrogram bin (line and both sidebands summed into one bin), a 45–2902 Hz scalogram axis on which the sideband span is 0.4 % of the height, a waveform at 33 samples per pixel, and a reshaped image covering 82 ms against a 181 ms modulation period. All four were re-parameterised (see _provenance.json → generation.perception_render); the slice also grew from 0.5 s to 8 s, because a segment shorter than ~2 s cannot resolve the sideband at any load.

Measured, not asserted. A small ViT (0.55 M params, patch 16, 40 epochs × 3 seeds) trained from scratch on the images alone, split by repetition exactly as shipped, each view fed at a size that preserves its own structure. Balanced accuracy on the held-out repetitions:

view	healthy vs faulted, old settings	new settings	1/2/3/4 bars
`spectrogram`	56 %	97 %	73 %
`reshaped`	51 %	77 %	44 %
`waveform`	93 %	86 %	63 %
`scalogram`	50 %	50 % (see below)	25 %

Chance is 50 % / 25 %. Three readings matter more than the headline numbers:

Use these for is there a fault, not for the bar count. Every view scores far lower on the 4-way task, and the best of them (73 %) merely matches what the sideband margin plus the load already give on their own (74 %) — the images add nothing for counting. See the answer-space note below.
waveform scoring lower is an improvement, not a regression. Under the old settings it reached 77 % on the bar count, above the 74 % the physics supports; with one physical rotor per class that excess can only be rotor recognition. Decimating to 2 kHz removed the shortcut (77 % → 63 %), which is what you want in a training corpus.
scalogram is 50 % under the ViT yet the evidence is measurably in the pixels. The CWT is constant-Q and cannot resolve 5.66 Hz at 60 Hz, so the modulation appears not as two lines but as the 60 Hz band pulsing ~45 times across the 8 s window. Detrended brightness in the sideband rows, FFT power at 2·s·f_line over its neighbours, healthy vs 4 broken bars at 4.0 N·m: 4.4 vs 42.7. Thresholding that one number blind on the test split gives 75 % overall and 92 % at 1.0 N·m. The band was chosen by this measurement and not by convention — 45–2902 Hz gives 1.9 vs 1.9 and the conventional 0.5×–3× band (30–180 Hz) gives 2.9 vs 3.0, both blind, while squeezing tighter than 40–80 Hz is worse (50–72 Hz: 2.7 vs 8.2). Any downstream resize below ~256 px wide aliases the pulsing away — that is why the ViT reads 50 % where a targeted measurement reads 75 %.

Two limits are unchanged by any of this. At 0.5 N·m the signal itself carries no evidence (healthy and faulted margins both sit near 35 dB), which is why 26 of that load's 50 runs are tiered absent — no rendering can recover what is not there. And because healthy is one physical rotor, any absolute accuracy on this corpus is an upper bound contaminated by rotor identity; the trustworthy comparison is old-vs-new, where that confound cancels.

Answer space. annot is the five-way gold (healthy … 4_broken_bars) so the count is not lost, but the evidence supports only fault present vs healthy — adjacent bar counts do not separate at any load, on the signal or in these images. Train and report binary; treat the count as metadata. Same shape as the SEU gearset release.

The representation that carries the physics for reasoning is the zoomed dB spectrum in AI4Manufacturing/IEEE-BRB.

Rig

WEG W22 three-phase induction motor, 1 cv, 220/380 V, 4 poles, 60 Hz, 4.1 N.m and 1715 rpm nominal, squirrel-cage rotor of 34 bars, coupled to a DC generator that sets the load torque. Faults are physical: the rotor was drilled with a 6 mm bit (wider than a bar) at mid-length, and five rotors were tested — healthy, then 1, 2, 3 and 4 adjacent broken bars. Every run is a direct start from the mains with balanced three-phase supply. That detail matters: with no inverter there is no current controller regulating the rotor asymmetry away.

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 / 1_broken_bar / 2_broken_bars / 3_broken_bars / 4_broken_bars
`reasoning`	chain-of-thought (empty here; filled in the `IEEE-BRB-annotated` sibling)
`cate` / `task`	`C` / `T-C1` (signal fault classification)
`metadata`	JSON string: representation, `sideband_margin_db` and the per-line LSB/USB dB and Hz, slip + its source, offset in Hz and FFT bins, `resolvable`, the per-load `margin_threshold_db`, computed_verdict, evidence_tier, load_nm, repetition, line_hz, fs, file, image_sha256, split

Splits

train / test = repetitions 1-7 / 8-10 (leakage-safe).

Provenance

Generated deterministically by forge_agent/examples/ieee_brb/convert.py (6ceba43daf) → forge_model/IEEE_BRB/convert_ieee_brb.py (6e138e37d4); see provenance.json for the full record.

How the evidence is measured

A broken bar makes the rotor magnetically asymmetric, which modulates the stator current and puts a pair of lines at f_line·(1 ± 2ks). The measured scalar is how far below the supply line the stronger of them sits (sideband_margin_db; smaller = stronger fault).

Two rules make that measurement honest, and both are there because the first build got them wrong:

The sideband position is predicted, never searched for. Allowed to roam, a peak search finds a "sideband" in perfectly healthy data — it locks onto the supply line's own leakage skirt. Measured at 4.0 N.m, a free search read healthy 39.8 dB and 1-broken-bar 40.2 dB (indistinguishable, both implying an impossible 0.57 % slip); anchoring on the nameplate slip instead gave 47.0 vs 41.2 dB. The slip comes from registry.get_motor() scaled by the recorded load — an operating condition, not the label.
The search window scales with the offset. The distance from the line to the first sideband is 2·s·f_line, which collapses from 5.5 Hz at full load to 0.69 Hz at 0.5 N.m. A fixed ±1.5 Hz window is then wider than the whole offset and swallows the supply line itself.

The gate

A run "shows fault evidence" when its margin falls below that load's threshold, and the threshold is mean − 3·std over that load's train-split healthy runs only:

load	threshold
0.5 N.m	34.0 dB
1.0 N.m	34.5 dB
1.5 N.m	36.5 dB
2.0 N.m	37.0 dB
2.5 N.m	38.0 dB
3.0 N.m	39.6 dB
3.5 N.m	42.7 dB
4.0 N.m	44.9 dB

This track keeps all runs (the tier grades the spectrum, not these images — see the card).

Caveats

The evidence gate is measured on the SIGNAL, not on the image. mcsa_sidebands reads the raw phase current; it never opens the PNG. So evidence_tier says this signal supports the label — not this image shows it. Every record carries evidence_measured_on: "signal". Gating this way is deliberate: the computation is deterministic and will give the same answer in a year, whereas filtering by what a model can already read would select for what it already knows. It is sound but incomplete — the image is a lossy function of the signal, so no-evidence-in-signal does imply no-evidence-in-image, but a record can pass and still lose its evidence in rendering. Image-side readability has never been measured. When it is, the rule is: evidence missing from the signal → drop the record; evidence present but unreadable in the image → change the rendering. This dataset already follows the second half — the zoomed dB spectrum exists because the unzoomed one hid the sidebands entirely, and the four perception renders are documented as unable to show the fault rather than dropped.
The fault is only visible under load. This is the dataset's defining property, measured across all 400 runs. At 3.0 N.m and above all four fault levels separate from healthy with no overlap; at 0.5 N.m nothing separates, because a lightly loaded motor barely slips and the sidebands crowd the supply line. Nothing hard-codes a load cutoff: the per-load evidence gate simply finds less and less to confirm, and the reasoning track thins out accordingly (confirmed runs by load: 0.5 N.m 23/50, 1.0 37/50, 1.5 42/50, 2.0-3.0 48/50, 3.5-4.0 50/50).
The marked sideband positions are PREDICTED, and the real peaks sit slightly closer to the supply line. The nameplate slip runs ~15-25 % high, so the measured peak is 0.19 Hz (0.5 N.m) to 0.78 Hz (4.0 N.m) inside the marker. The peak search window absorbs that at every load, but at 0.5 N.m the peak lands 92 % of the way to the window edge — which is the mechanical reason the light loads are fragile. A per-load slip calibration would remove the bias and is the obvious next improvement; this build ships the nameplate slip and states the offset.
"Separable" has two senses here and only one holds. Each fault class against healthy separates with no overlap at 3.0/3.5/4.0 N.m (4 of 4 classes). Adjacent fault classes do not separate at any load: the best cases, 3.5 and 4.0 N.m, reach 3 of 4, and 3.0 N.m only 2 of 4. At 4.0 N.m one of the ten 4-broken-bar runs reads 44.0 dB against 29-31 dB for its siblings — a healthy-looking outlier that happens to sit in the test split. So no record's evidence supports the bar count; every confirmed record supports only "a fault is present".
The verdict is binary; the gold is the bar count. The margin is monotone in the count (measured at 4.0 N.m: healthy 45–49 dB, 1 bar 40–42, 2 bars 34–37, 3 bars 32–33, 4 bars 29–31), but the detector is only asked to attest that evidence is present — grading to an exact count would need four label-derived thresholds. Same shape as the SEU gearset release.
Per-load calibration. A healthy rotor's margin depends strongly on load, so the threshold is computed per load group from that load's train-split healthy runs only (mean − 3·std) and applied blind. No threshold was adjusted against the faulted rotors.
No windowing. One 20.02 s run is one record. MCSA resolution is the whole game: the full record gives 0.05 Hz bins, and halving the window would halve that — at light load the sidebands are already only ~14 bins from the supply line.
Split — train/test follow repetitions 1-7 / 8-10. A run is never split, so it is leakage-safe. A rotor-wise split would be the strong one and is not available: each class is exactly one physical rotor, and all 8 loads of that rotor sit in the data, so a model may learn the rotor rather than the fault. Same shape of limitation as the SEU release (one specimen per class/condition).
Adjacent rotors. The broken bars are always adjacent, which is the easier case; scattered breaks produce weaker sidebands and are not represented here.
Two points where the source page and the data disagree, both recorded rather than guessed. The page says each run is 18 s; the files measure 1,001,000 samples at 50 kHz = 20.02 s, and the data is what is used. The page describes the eight load levels as 12.5–100 % of full load (0.5125–4.1 N·m) while the in-file group names torque05–torque40 are read as 0.5–4.0 N·m; at full load that 4.0-vs-4.1 gap is a 2.4 % slip difference = 0.13 Hz in the predicted sideband position, against a ±1.5 Hz search window, so no verdict changes. The upstream readme.txt would settle it and has not been retrieved.

Source & license

Source: Experimental database for detecting and diagnosing rotor broken bar in a three-phase induction motor, Laboratory of Intelligent Automation of Processes and Systems (LAIPS) and Laboratory of Intelligent Control of Electrical Machines (LACIME), School of Engineering of São Carlos, University of São Paulo, Brazil — IEEE DataPort, doi:10.21227/fmnm-bn95.

License: CC BY 3.0 US (Attribution). Checked 2026-08-05: the dataset page itself names no licence; the licence comes from IEEE DataPort's Terms of Use, which place uploaded content under the Creative Commons Attribution (CC BY) License, 3.0 US. CC BY permits redistribution and derivative works with attribution, so this derived dataset is redistributed under the same terms.

Please cite the source dataset:

Aline Elly Treml, Rogério Andrade Flauzino, Marcelo Suetake, Narco Afonso Ravazzoli Maciejewski, Experimental database for detecting and diagnosing rotor broken bar in a three-phase induction motor., IEEE Dataport, September 16, 2020, doi:10.21227/fmnm-bn95

Downloads last month: -

Total file size:

56.9 MB