AI4Manufacturing/DUALDISCO-perception

DUAL DISCO — perception representations (grounding track)

Part of the AI4Manufacturing FORGE corpus (Category C, task T-C1). Four image renderings of the same airborne acoustic recording from a dual-laser LPBF machine, one HF config each.

Rows: 4480 across 4 configs (1120 each — 560 recordings × 2 microphones); labels {'conduction': 400, 'keyhole': 720}; shipped unsplit.

config	image	how much signal it shows
spectrogram	short-time Fourier transform, 0–100 kHz	0.2 s
scalogram	Morlet continuous-wavelet transform, 1–95 kHz	20 ms
waveform	the raw time trace, anti-alias decimated	0.2 s
reshaped	samples folded into a 64×64 grayscale grid	20 ms

from datasets import load_dataset
ds = load_dataset("AI4Manufacturing/DUALDISCO-perception", "spectrogram")

The reasoning counterpart — the averaged spectrum, where the answer is a position on the frequency axis — lives in AI4Manufacturing/DUALDISCO.

⚠️ These four views show a 0.2 s window, not the whole melt

The window is taken a quarter of the way into the stretch where the laser trigger is high; measured across records its RMS is 1.08–1.57× the whole stretch's, so it always lands on the process rather than in a pause. But the scalogram and reshaped views cover only 20 ms — the wavelet transform resamples above 4,096 samples, and a 64×64 grid is 4,096 samples — so at 200 kHz they are a much narrower view than the same two renderings give on a 12 kHz bearing record. The reasoning track's spectrum, by contrast, averages the whole engaged stretch.

reshaped uses a fixed amplitude scale for the whole dataset (the 99.9th percentile of the calibration side), not a per-image stretch: otherwise loudness would be normalised away image by image.

No evidence gate is applied on this track: every row ships. metadata.evidence_tier is informational here — nothing was dropped or reordered by it.

Rig

DUAL DISCO: a laser powder-bed fusion machine with two lasers, each melting its own cube on the same build plate, and two condenser microphones listening in the air around the build volume. Four channels at 200 kHz — a laser trigger, the two microphones, and a fourth continuous channel the release does not name — 1,700,000 samples (8.5 s) per recording.

Each recording is labelled with the melt-pool mode of both cubes: Conduction (a shallow, wide pool) or Keyhole (the beam drills a vapour cavity and the pool goes deep). The mode is set by the laser parameters; a numbered laser setting (1–4) accompanies it. The two lasers run either Sequential (one cube then the other) or Simultaneous (both at once).

Why the reasoning track measures a frequency

Nothing in this machine rotates, so there is no characteristic frequency to compute and no line to look for — the whole apparatus the corpus's bearing datasets lean on does not apply. What changes with the melt pool is the shape of a broadband spectrum: a keyhole is a deeper, slower-oscillating cavity and it moves the sound's energy down in frequency.

So the judging number is a frequency, not an amplitude: f_q25, the frequency below which a quarter of the energy above 1 kHz lies. A microphone's sensitivity divides out of it, and it states the direction of the shift instead of hiding it inside the sign of a ratio.

One recording, two cubes — and why the question is binary

The source labels both cubes of every recording, and in 240 of the 560 they differ — one cube in conduction, the other in keyhole, at the same time, heard by one pair of microphones that cannot say which sound came from which cube. One recording, two answers.

So the question here is binary: was ANY laser in keyhole mode. Those recordings are keyhole. That is also the question the process actually asks — a keyhole traps vapour and leaves pores in the part, and which cube did it is a second question this recording cannot answer.

Every query in this repo says so in words: it names the two lasers and spells out that keyhole means at least one of them, conduction that neither was. The queries do not say which way the spectrum moves — that is the thing being asked.

recording	count	annot
both cubes conduction	200	conduction
one cube each	240	keyhole
both cubes keyhole	120	keyhole

The folded-in recordings are not noise, and that was checked before folding them in. A recording with one cube per mode should sit BETWEEN the two pure classes, and it does:

	pure conduction	one cube each	pure keyhole	AUC cond vs mixed	AUC mixed vs key
mic1	17.46 kHz	15.98 kHz	14.98 kHz	1.000	0.937
mic2	19.74 kHz	16.80 kHz	15.28 kHz	1.000	0.899

metadata.both_cubes_same_mode, cube1_mode and cube2_mode travel with every row, so the 640 unambiguous rows are one filter away if you want conduction-vs-keyhole instead of any-keyhole.

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: conduction / keyhole
reasoning	chain-of-thought (empty here; filled in the -annotated sibling)
cate / task	C / T-C1 (signal classification)
metadata	JSON string: record_uid, mic, operation, laser_setting, layer, prelude_layer, cube1_mode, cube2_mode, both_cubes_same_mode, upstream_archive, upstream_split, fs, engaged_s, engaged_frac, gate_threshold_v, analysis_band_hz, band_ratio_band_hz, nper, f_q25_hz, f_q50_hz, f_q75_hz, centroid_hz, spread_hz, flatness_db, band_ratio_20k_100k, rms, kurtosis, window_rms_ratio, decision_feature, decision_threshold_hz, decision_margin_hz, calibration_row, computed_verdict, evidence_tier, evidence_measured_on, representation, image_sha256

Splits

Shipped unsplit — there is no defensible train/test division to give you.

Every recording comes from one machine, one build plate and the same two microphones. There is no held-out unit to move to a test set, so no division of these rows puts unseen hardware on the far side.

Note on the train split name. HuggingFace requires every split to be named, so the single unsplit set is served as train. It is the whole dataset, not a training portion — there is no matching test. It returns all 1120 rows.

Two axes are available if you need one, and they are not equivalent.

• metadata.laser_setting (1–4) is the closest thing to a real generalisation test here: setting 4 appears in only one of the two upstream archives, so training on 1–3 and testing on 4 asks whether the criterion survives a parameter set it has never seen.
• metadata.upstream_split records the source's own train/test division, which is by laser setting. This release does not ship it as a split — it is used for one thing only: the decision threshold was calibrated on the train archive's microphone 1 and nothing else.

Whatever you pick, split on metadata.record_uid, never on rows. Each recording contributes two rows, one per microphone, and they are the same physical event heard twice. This repo holds 1120 rows from 560 recordings; a random row split puts the same melt pool on both sides.

Provenance & the evidence gate

Generated deterministically by forge_agent/examples/dualdisco/convert.py (49e9efe4a2) → forge_model/DUALDISCO/convert_dualdisco.py (b47aa516bb); full record in provenance.json.

Labels come from the source's own label sheets. An evidence gate — a label-independent computation on the raw sound — decides what may enter the reasoning track: f_q25 above the threshold reads conduction, below it keyhole.

There is one threshold per microphone, and both are read off one archive only.

	threshold	empty interval	calibration rows
mic1	16.67 kHz	121 Hz	400
mic2	18.16 kHz	512 Hz	400

Why not one number for both: a recording with one cube per mode sits between the two pure classes, so folding it into keyhole raises that class's ceiling until the empty interval on microphone 1 is only tens of hertz — far thinner than the ~2 kHz offset between the two microphones. Measured on all rows, both fitted on the calibration archive only: one global threshold 1003/1120, one threshold per microphone 1120/1120. A number per sensor is an ordinary calibration; pretending one number covers both microphones would not be.

Every row outside the main archive is scored blind against those thresholds: 320/320 (calibration archive 800/800). The search never sees the label.

The engaged stretch is cut with the recorded laser trigger, not an inferred threshold: cut_segment(gate=...) takes the midpoint between the trigger channel's two voltage levels. Which channel is the trigger is decided from the signal's shape and then checked against Settings_0.txt ("Trigger Source: Analog Input 0"); a disagreement raises.

On this track the gate is none (labels are the source's own; evidence rides along as metadata).

Caveats

• Two rows per recording, one per microphone — split on record_uid. They are the same event heard twice. The two microphones are not interchangeable in absolute terms: a threshold fitted on one and carried to the other, unchanged, gets 430/560 on f_q25 and 242/560 on the band ratio. So this release ships one threshold per microphone, and metadata.mic says which one a row belongs to. A model trained here is allowed to use that; a model deployed on a third microphone would have to calibrate again.
• The label does not say WHICH laser. 480 rows had one cube in each mode and are labelled keyhole because the question is is any laser in keyhole. Nothing in this repo can answer which cube — the two lasers fired into one pair of microphones. If you need the unambiguous three-way version, filter on metadata.both_cubes_same_mode.
• The classes are unbalanced: {'conduction': 400, 'keyhole': 720}. Report per-class numbers, not accuracy.
• 160 rows come from 'Prelude' layers, and every one of them is conduction at laser setting 2. A layer name that predicts the label perfectly is exactly the shape of a leak, so it was checked: prelude vs body-layer conduction at the same setting separates at AUC 0.558 (mic 1) / 0.357 (mic 2) — i.e. they are not distinguishable, and they are kept. metadata.prelude_layer carries the flag so you can drop them if you disagree.
• The firing mode is audible, and it is not the label. Within one class, f_q25 tells Sequential from Simultaneous at AUC 0.951 / 0.823 (mic 1, conduction / keyhole). The two modes are balanced across the classes, so this cannot leak the answer — but it is real variance inside each class, and a small split that happens to be unbalanced on metadata.operation will measure firing mode. Stratify on it.
• The label is a commanded regime, not an observed pool. The source ships parameter sheets naming each cube's mode; no cross-section, no CT, no high-speed imaging travels with the release. Take annot as the machine was set to melt this way.
• The analysed stretch is the whole time the laser trigger is high (median 75% of the recording), which includes the quiet moments between bursts. Tightening it to the loud parts would be defensible; it was not needed, and the looser definition is the one the numbers above were measured on.
• The fourth channel is unidentified. It is continuous like the microphones but ~60× quieter, and the release does not say what it is. It is not used here and not shipped as an image; the adapter returns it as signal_low_level rather than guessing.
• 560 recordings is small, and it is one machine. The value is the kind of evidence — a broadband spectrum whose position, not whose level, carries the answer — not the volume.

Source & license

Source: DUAL DISCO — acoustic emission from a dual-laser LPBF process. Data: doi:10.5281/zenodo.13863838, CC BY 4.0. The package holds Raw_data.zip / Raw_data_test.zip (parquet, one file per channel per recording), the label sheets params.xlsx / params_test.xlsx, and the acquisition settings Settings_*.txt.