SNU-PLANETARY-annotated

Order spectra about the gear mesh line of a single-stage planetary gearbox (ring fixed; sun 31 teeth, 3 planets of 31, ring 95), as a three-class fault location task: normal / sun_gear / planet_gear. 917 records. reasoning is filled on every record; the twin repo SNU-PLANETARY is identical except that field is empty.

The reading

The abscissa is the order offset from the mesh line; the ordinate is every line's height divided by the mesh line, so the mesh line itself sits at 1 and the axis carries no information about which record you are looking at. Two marks decide it, both drawn on every record including the healthy ones:

step	read	verdict
1	first carrier sideband left of the mesh line	>= 1.42 -> planet_gear
2	first carrier sideband right of it	>= 0.38 -> sun_gear, else normal

Measured ranges over the 917 shipped records:

class	step-1 quantity	step-2 quantity
normal	0.92 – 1.11	0.10 – 0.36
sun_gear	0.86 – 1.18	0.40 – 1.20
planet_gear	1.70 – 3.71	—

Both steps separate with no overlap, and the criterion agrees with the gold on 917 of 917.

Why left and right differ. The three planets travel. A fault carried on one swings past the fixed accelerometer once per carrier revolution, so the path from the damage to the sensor is modulated at the carrier rate and the mesh line grows carrier-spaced sidebands -- the classic planetary result (McFadden, JSV 144 (1991) 163-172; Inalpolat & Kahraman, JSV 323 (2009) 677-696). The sun does not orbit; it turns in the middle and rolls past each planet in turn, so it gets no such near/far modulation and its mark lands on the other side.

The honest number

The shipped split is the conventional one -- per configuration, the first 80% of windows train, the last 20% test -- and it leaks the assembly: both halves come from one build. The number that says something is leave-one-assembly-out, where a whole build is held out and the thresholds are fitted on the others only:

assembly	what it is	gold	thresholds	refitted	right
A1	Normal	normal	1.42 / 0.38	step 2 falls back	131 / 131
A2	Sun-0.50	sun_gear	1.42 / 0.54	both refitted	130 / 131
A3	Sun-0.75	sun_gear	1.37 / 0.38	both refitted	131 / 131
A4	PR-0.75	planet_gear	1.42 / 0.38	both refitted	131 / 131
A5	PR-1.00	planet_gear	1.53 / 0.38	both refitted	131 / 131
A6	PR-1.25	planet_gear	1.42 / 0.38	both refitted	131 / 131
A7	PS-0.75	planet_gear	1.42 / 0.38	both refitted	131 / 131
all					916 / 917

Across the folds that could refit it, the step-1 threshold moves only between 1.37 and 1.53, so it is not tuned to any one build. One fold cannot refit everything: A1 (step 2 falls back). A threshold here is the midpoint between the two sides it separates, and holding out the single healthy build leaves step 2 with no healthy side to measure against, so that fold reuses the corpus value. Its 131 records are therefore a weaker test than the other six folds, not an independent one.

What this set does not claim

• Severity is not a label. The sun's two groove sizes are monotonic in the step-2 quantity but the planet-ring's three are not, so only location ships.
• The healthy class is one build. Planet faults span four separate assemblies and sun faults two; normal was recorded once, five weeks before the rest. The sun_gear vs planet_gear edge therefore rests on more evidence than the normal vs faulted edge.
• The variable-speed recordings are not here. The same criterion does not survive the sinusoidal sweep (521-1588 rpm): restricted to windows inside the stationary speed band the sun/normal step returns but the planet step stays scattered. Those 609 records are in SNU-PLANETARY-perception, where the label comes off the rig instead of off this criterion; they are not in this track.
• One accelerometer. No vibration-separation or per-tooth work is possible here; that needs a ring-gear array.

Fields

query (30 phrasings) · image · annot (one of normal, sun_gear, planet_gear) · reasoning (the chain of thought, closing on a line FINAL ANSWER: <class>) · cate C · task T-C1 · metadata (tooth counts, orders, the measured sideband heights, assembly id, window index, rpm, style variant, split).

Splits: test 189, train 728.

Figures

11 drawing styles are used across the set -- colours, canvas, ticks, grid, spines, marker as dashed line or band, label orientation, font. What never varies: the curve, the plotted band, which marks are drawn, and the ordinate limit, which is pinned for the whole corpus so the scale cannot become a shortcut.

Source

SNU Planetary Gearbox Dataset, Yoo, Kim, Ha & Youn, figshare 10.6084/m9.figshare.32025606 (v5), CC BY 4.0. Vibration PCB 601A02, encoder ONO SOKKI MP-981 with a 60-tooth wheel, 25,600 Hz. Cite the upstream dataset.