AI4Manufacturing/PHM2010-perception

PHM 2010 milling — perception tracks

Part of the AI4Manufacturing FORGE corpus (Category C, task T-C1). Four HF configs, one per representation of the same axial cutting force: spectrogram, scalogram, waveform, reshaped. The question is a plain classification — worn or serviceable — with no rule stated, because a newton threshold cannot be applied to a spectrogram. The rule-bearing representation lives in the PHM2010 repo.

config	records	splits	labels
spectrogram	945	{'train': 630, 'test': 315}	{'serviceable': 657, 'worn': 288}
scalogram	945	{'train': 630, 'test': 315}	{'serviceable': 657, 'worn': 288}
waveform	945	{'train': 630, 'test': 315}	{'serviceable': 657, 'worn': 288}
reshaped	945	{'train': 630, 'test': 315}	{'serviceable': 657, 'worn': 288}

Unfiltered. Unlike the reasoning track these configs keep every cut, including the ones where the reasoning track's rule disagrees with gold. Those disagreements are the honest difficulty of the task, not defective records.

Every renderer default was recalibrated. They were set on 12 kHz bearing data; this rig samples at 50 kHz and its diagnostic content sits below 1.8 kHz, so four defaults meant something different here. The reshaped render is additionally cycle-locked — one row is one spindle revolution — which the bearing family cannot do because its rigs run at more than one speed. Exact parameters and the reason for each are in provenance.json under render_calibration.

Rig

Process	dry high-speed milling of stainless steel HRC52, down-milling
Cutter	3-flute tungsten-carbide ball-nose end mill, 6 mm diameter
Cut	0.2 mm axial depth, 0.125 mm radial depth, 10400 rpm, 1555 mm/min feed
Sensors	Kistler 3-axis dynamometer + 3-axis accelerometer + acoustic-emission RMS
Sampling	50 kHz, 7 channels; a cut is 2.5-5 s
Population	3 instrumented cutters x 315 cuts = 945 records
Wear label	flank wear VB per flute, read offline on a LEICA MZ12 microscope

Schema (7-field unified record)

field	here
query	the paraphrased question; for the reasoning track it also states the decision rule
image	the rendered PNG, bytes embedded
annot	worn or serviceable
reasoning	None — filled by the annotation pass, not here
cate / task	C / T-C1 (signal fault classification)
metadata	JSON: cutter, cut index, measured period and rpm, wear in um (all three flutes), the measured peak-to-valley and the threshold it was compared with, split

Splits

train / test = by cutter, never by cut.

Provenance & calibration

Produced by forge_agent/examples/phm2010_milling/convert.py (forge_agent e8726e30f2) and packaged by forge_model/PHM2010/convert_phm2010.py (forge_model 7985ec9066).

Gold. worn = mean flank wear >= 125.0 um. Chosen by us, from label-only properties; see caveat 3.

The measurement in the query. Peak-to-valley of the angle-folded axial force. The threshold, 74.5 N, is fitted on the TRAIN cutters (c1, c4) only and applied blind to c6 — the same discipline as this corpus's other calibrated gates. Fitting it across all three cutters instead reaches a flattering number by fitting on the test set; both are recorded in provenance.json.

cutter	cuts	rule agrees with gold	majority baseline	missed worn	false worn
c1	315	93%	71%	22	0	train
c4	315	97%	78%	0	11	train
c6	315	83%	59%	54	0	held out

The threshold does not transfer cleanly between cutters — per leave-one-out fold it lands at different values, because the three cutters' healthy baselines differ about fourfold. That is a property of the rig, and it is why the rule is absolute newtons rather than a ratio to a cutter's own first cut: at equal wear the ratios spread 3.5x while the absolute forces spread 1.5x.

Query pool. Domain cutting_force, angle_profile 30, spectrogram 30, scalogram 30, waveform 30, reshaped 30 paraphrases.

Caveats

1. The wear labels are interpolated, not measured per cut. The archive ships 2835 values and the source paper says each cut was read under a microscope. The values do not support that: none ever decreases, 231 repeat the previous cut exactly, and two of the nine flutes are reproduced to 0.0017 um by piecewise-linear interpolation through 15 cut indices. The test that settles it is not monotonicity — NASA Ames' genuinely per-cut-measured milling set is monotone too, at a higher rank correlation — it is predictability: hide a value and predict it from its neighbours, which costs 0.0017 um here and 12.5 um there, a factor of 7494. A microscope does not read to a thousandth of a micrometre.
2. Which is why the split is by cutter and the cut index never appears in an image. On a random per-cut split a nearest-neighbour model that reads no sensor at all matches the published deep models. Nothing rendered here encodes how far into a cutter's life a record sits.
3. The wear threshold is ours. The archive contains no threshold, no pass/fail column and no documentation — it was a regression challenge scored in micrometres. ISO 8688-2's 300 um end-milling criterion is never reached (the run stopped at 315 cuts; worst flute 234.7 um). We chose the limit from label-only properties, not from any detector's accuracy.
4. Three cutters is at most three folds. One is held out. This says nothing about transfer to another machine, another workpiece or another cutter geometry.
5. The task is serviceability, not wear regression. Predicting the micrometre value on this dataset trains the interpolation formula back.

Source & license

PHM Society 2010 Data Challenge, milling tool wear. Rig and protocol: Li X, Lim B S, Zhou J H et al., Mechanical Systems and Signal Processing 23(8), 2009, doi:10.1016/j.ymssp.2009.06.008. The official distribution host has lapsed; the archive used here is a byte-identical mirror, verified against four independent copies of the label files.