IEEE-BRB-annotated

Stator-current spectra of a mains-fed induction motor around the supply line, as a binary rotor-cage task: healthy / broken_bars. 477 records. reasoning is filled on every record; the twin repo IEEE-BRB is identical except that field is empty.

The reading

One quantity, and it is what an eye judges on a spectrum: at the pair of marks either side of the supply line, how far does the line stand above its own two shoulders.

Measured over the 477 shipped records, per load, as a ratio to what sits immediately beside the mark:

load	threshold	healthy	broken_bars	gap
2.5 N·m	1.33x	1.01 – 1.25x	1.87 – 4.63x	1.50x
3.0 N·m	1.47x	1.05 – 1.32x	2.09 – 6.19x	1.59x
3.5 N·m	2.02x	1.11 – 1.75x	3.94 – 11.90x	2.24x
4.0 N·m	2.35x	1.42 – 2.24x	6.12 – 19.12x	2.73x

The line is fitted on healthy training records only -- no faulted record and no test record touches it -- as the mean plus three standard deviations of the healthy training prominence at that load. On the shipped set it is right on 477 of 477.

Why the line moves with load. More load means more slip, which pushes the sidebands further out and makes them easier to separate; but it also brings out the sound rotor's own unevenness, which no real cage is without. So the same small bump counts at 2.5 N.m and does not at 4.0. The load is in the query for that reason; it is an operating condition, every load carries both classes, and it says nothing about the answer on its own.

Where the marks come from. The supply frequency is read off the spectrum and the shaft speed off the shaft line in the frame accelerometer -- a channel that never sees the current sideband. Together they fix the slip, and the marks go at twice the slip either side of the supply line. The previous release took the slip from the nameplate scaled by load, which runs about 20 % high: at 4.0 N.m its mark sat 0.78 Hz outside the peak it was pointing at.

The honest number

leave one repetition out -- the line refitted on the rest each time:

held out	right
each of 10 groups	477 / 477

leave one phase clamp out -- the line refitted on the rest each time:

held out	right
Ia	156 / 159
Ib	159 / 159
Ic	159 / 159
all	474 / 477

leave one faulted rotor out -- the line refitted on the rest each time:

held out	right
r2b	120 / 120
r3b	120 / 120
r4b	117 / 117
rs	0 / 0
all	357 / 357

120 records fell in a fold whose line could not be refitted and are not counted.

A load cannot be held out, because the line is fitted per load and holding one out leaves nothing to fit it with. What the rotor fold asks is the useful question: the line is placed by healthy records alone, so an unseen faulted rotor is genuinely unseen.

The three clamps are not quite interchangeable. Every miss in the phase fold is the same thing: three healthy runs on Ia at 4.0 N·m, where a line fitted on Ib and Ic alone comes out about 1.3 dB too tight. At that load the sound rotor's own residual line is a little taller on Ia than on the other two clamps, so a threshold that never saw Ia clips it. Worth knowing before pooling the three.

What this set does not claim

• The healthy class is one physical rotor. So is each faulted class. Any accuracy on this data is an upper bound contaminated by rotor identity, and that is why only the binary task ships.
• One broken bar is not in this track. Read the same way on all three clamps, it overlaps the healthy rotor at 2.5 N.m. From 3.0 to 4.0 N.m every one-bar run clears the healthy maximum, but the lowest of them only by 1 to 12 %, and the per-load line would still call 13 of 30 runs at 3.5 N.m healthy. A margin that small cannot be seen on the figure -- the known limit of current-signature analysis on a small machine, not a defect in the recording. Those runs are in the perception track.
• Light loads are not in this track. Below 2.5 N.m the marks fall inside the supply line's own skirt or the gap shrinks to about a decibel. 4 load settings ship.
• Three phases of one run are three views, not three samples. All three clamps are read and all three ship, with channel in metadata; the split is by repetition, so a run never straddles it.
• How many bars is metadata, never the answer. broken_bars (2, 3 or 4) is recorded and is deliberately not the label.

How reasoning was produced

This revision (2026-09-25) replaces the reasoning column; earlier versions remain in the repository history.

A deterministic renderer, not a language model. Each text is assembled by fixed code from the record's metadata and its gold label. No API, no sampling, no model output: rendering the same record again gives the same bytes.

Four paragraphs, separated by blank lines, with no headings inside the text:

1. Figure reading — what the figure is (the current spectrum of one stator phase from 50 to 70 Hz, in dB relative to the supply line, with marks at the supply frequency and at (1-2s)f and (1+2s)f), the slip and the load of this record, and how far each sideband's peak stands above its own shoulders, in whole dB. Description only: no explanation, no verdict.
2. Physics — both candidate classes explained in one continuous derivation: a squirrel cage, the slip and the slip-frequency bar currents; why an intact cage puts nothing but the supply line into the stator current; how a broken bar leaves a hole in the current pattern that splits into a forward and a backward wave, the backward one giving the lower sideband; how the resulting torque pulsation makes the speed wobble and adds the upper sideband, whose size inertia decouples from the lower one; why a sound cage still shows small traces; and why more load means more slip, larger bar currents and taller sidebands for the same rotor. Every term is defined where it is first used. The content of this paragraph does not depend on the record's label.
3. Linking — the height of the taller sideband is weighed for the record's own load. Because the same rotor shows taller sidebands at higher load, a given height says more at a light load than near rated load; the paragraph says, in words, whether this height at this load is the small trace a sound cage leaves or the rise a broken bar produces, and whether both marks carry a peak.
4. Conclusion — one sentence, then a blank line and FINAL ANSWER: <annot>, equal to the record's gold label byte for byte.

Variation comes from the image, never from the label. Paragraphs 1 and 2 swap order for roughly half the records (the exact share is given below); the linking paragraph and the conclusion always come last. Every sentence has at least three phrasings, chosen sentence by sentence. The swap and every choice of phrasing are taken from the SHA-256 of the record's image bytes (the image column; metadata has no image hash), never from the label.

No criterion in the text. No decision threshold, class boundary or per-class range appears anywhere in the text, and no reading is compared against one. Sideband heights are the record's own, rounded down to whole dB; at a given load the same printed height never occurs under both labels.

Numbers for this set. 477 records; median length 1014 words (whole text, FINAL ANSWER line included); the physics paragraph comes first in 217 of 477 (45.5%). All 477 texts are distinct, and every text ends on its gold label (checked on all rows).

Example. The linking paragraph (paragraph 3) of test record 76 (0-based row of the test split, gold broken_bars), verbatim:

In this spectrum, the taller sideband is the upper one, about 9 dB above its shoulders, and the lower one rises too, about 8 dB, so both marks carry a peak. This is 2.5 N·m, about three fifths of rated load: with rotor currents this moderate, the slight unevenness of an intact cage hardly lifts the sideband positions. Because the same rotor shows higher sidebands the heavier it is loaded, a height read here says more about a broken bar than it would near rated load. A rise of about 9 dB is clearly more than a sound cage's slight unevenness produces at this load; it takes current forced round a broken bar, and the backward wave that creates, to raise a peak like this. Both sidebands being present fits the chain: the lower from the backward wave itself, the upper from the speed ripple it drives.

Fields

query (119 phrasings) · image · annot (one of healthy, broken_bars) · reasoning (filled) · cate C · task T-C1 · metadata (the measured shaft rate and slip, the sideband offset, both sidebands' depth and prominence, the load, the channel, the rotor tag, the bar count, the threshold in force, the style variant, the split).

Splits: test 141, train 336.

Figures

11 drawing styles are used across the set -- colours, canvas, ticks, grid, spines, fill, marks as dashed lines or as narrow bands, label placement, font. What never varies: the curve, the plotted band of ten hertz either side of the supply line, which marks are drawn (the supply line and the first-order pair, the three the criterion reads), and the ordinate limits, pinned for the whole corpus. No threshold bar and no noise-floor line is ever drawn: the floor is the grass the reader can see, and drawing it would put half the criterion onto the evidence.

Source

IEEE broken-rotor-bar dataset, WEG W22 1CV four-pole induction motor, 50 kHz current and 7.6 kHz vibration, 20.02 s per run. Cite the upstream dataset.