ITSC-annotated

Park-vector current loci of a three-phase induction motor, as a four-class stator winding task: normal / phase_A / phase_B / phase_C. 183 records. reasoning is filled on every record; the twin repo ITSC is identical except that field is empty.

The reading

Both coordinates are divided by the radius of the equal-area circle, so that circle is the 1.0 ring on every image and the tick numbers carry no current amplitude at all. Two steps, and both are drawn on every record including the healthy ones:

step	read	verdict
1	the long axis of the fitted ellipse against the short one	at least 1.09x longer -> a winding is shorted
2	which pair of A/B/C marks the long axis lines up with	that winding

Measured over the 183 shipped records:

class	long axis / short axis	angle to its own pair of marks
normal	1.034 – 1.084	—
phase_A	1.218 – 1.744	0 – 11 deg
phase_B	1.205 – 1.991	0 – 27 deg
phase_C	1.113 – 1.899	0 – 16 deg

Healthy and faulted do not overlap on step 1 (1.034 – 1.084 against 1.113 – 1.991), and on step 2 the nearest wrong pair of marks is never closer than 9 degrees further round.

Where the marks come from, and why they moved. A short unbalances one phase; the imbalance splits into a forward- and a backward-rotating part, and the sum of the two traces an ellipse whose long axis lies at half the difference of their phases. A third of a turn between windings therefore becomes about a sixth of a turn between answers -- the three seats sit roughly 60 degrees apart, not 120. Where that set of three lands on the dial depends on the machine. The previous release drew the marks at the winding angles 0/120/240 instead, and reading the long axis against them names the right winding on 62 of 168 faulted records: each class lands nearest the wrong mark. The marks are now at the measured seats -- A 140.03, B 89.46, C 27.45 degrees -- which is the same ruler the criterion uses. They are identical on every image and are stated in the card rather than hidden, because they were fitted from training labels.

The honest number

The rule speaks on all 195 candidate windows and is right on 183 of them (93.8%); the 183 it got right are what ship. That retention rate is the corpus's own precedent but it must not be read as an off-the-shelf detector's accuracy.

leave one recording out -- a whole group held out, the rule refitted on the rest:

held out	right	note
each of 61 recordings	183 / 183

leave one severity out -- a whole group held out, the rule refitted on the rest:

held out	right	note
0	0 / 15	no healthy record left to fit the shape step
10	36 / 36
20	42 / 42
30	45 / 45
40	45 / 45
all	168 / 183

The one fold that fails is the degenerate one. Holding out any single faulted severity leaves the rule untouched, because the step-1 threshold is a 3-sigma band around the healthy records and the faulted ones set nothing -- removing a severity does not move it. Holding out severity 0 removes the healthy class itself, and then there is nothing left to fit that band with at all; that fold is scored 0/15 above and it is a statement about the design of the split, not about the rule. Training on the two mild severities and testing on the two strong ones gives 90 / 90; the other way round, which is the harder direction, 77 / 78.

Leave-one-recording-out is the weak test here, not the strong one. A recording's three windows overlap, so holding one out removes near-duplicates rather than an unseen condition. The severity holdout above is the one that asks a real question, and there is no unseen-machine split on this data at all.

What this set does not claim

• The healthy class is five recordings. 15 windows in total, and the threshold stands on the 9 training ones. Every healthy-side number here moves in steps of about 17 percentage points.
• Severity is not a label. severity_pct (10/20/30/40 % of the winding, set on a switch panel) is in metadata only. It is close to monotone in the step-1 quantity, which is exactly why a model can learn it and why the severity fold above matters.
• One motor, one operating point. No load, no speed variation, 65 recordings. There is no unseen-machine split available on this data at all.
• The three windows of a recording overlap. 0.5 s at 0.25 s hop, so they are three views of one acquisition. The split is by repetition, so a recording never straddles it.

Fields

query (30 phrasings) · image · annot (one of normal, phase_A, phase_B, phase_C) · reasoning (filled) · cate C · task T-C1 · metadata (the two semi-axes and the equal-area radius, the axis angle, the angle to each pair of marks, the three measured seats, severity, the recording id, the repetition, the style variant, the split).

Splits: test 72, train 111.

Figures

11 drawing styles are used across the set -- colours, canvas, ticks, grid, spines, fill, locus as a line or as points, label placement, font. What never varies: the traced locus, the fitted ellipse, the equal-area circle, both axes of the ellipse, the three pairs of phase marks, and the axis limits, which are one constant for the whole corpus so the scale cannot become a shortcut.

Source

Inter-turn short-circuit dataset for three-phase induction motors (Cropped_Signals_SF), 1 kHz, 60 Hz supply, no load. The cropped release is used rather than the raw one because the raw recordings switch the short in mid-take. Cite the upstream dataset.