sankalpsthakur/well-production-event-causal

Well production-event detectors (Petrobras 3W)

Compact causal detectors for production events in real offshore well recordings from Petrobras 3W 2.0, tested on well groups they never saw during training. These are research models with their failures published. None is ready for deployment.

Results on unseen well groups (initial confirmation, scored 2026-10-04 UTC)

The validation-selected primary is a 36-parameter logistic model. Two causal TCNs have 4,961 parameters each.

Measure	Result	Budget
Episodes covered	24 of 55
Known onsets caught within 300 s	3 of 12
False alarms	42 over 9.124 normal days	1 per day
Normal time spent in alarm	64.392%	1%
Normal time with usable inputs	59.812%

Twenty-one of the covered episodes already had an alarm running when they began. In a later 15-method validation comparison (published 2026-10-05 UTC), the selected logistic model caught 13 of 24 onsets within 300 s, 11 of them from one recording group. IBM's pretrained TSPulse caught none.

What's here

• Every trained candidate and reference method, with the full report and deployment assessment
• The 15-method validation comparison and the TSPulse causal preflight
• Streaming replay and timing: median 3.46 and 3.35 ms per packet for the two TCNs on the observed Linux CPU, excluding file decoding and acquisition. This is not a target-device latency.
• Audit tables in the 3W audit dataset, and a live explorer

Reproduce

From the repository root, at a fixed revision:

uv run --no-project --python 3.12 --with numpy==2.2.6 --with torch==2.8.0 python tools/replay_well_checkpoints.py --artifacts artifacts --receipt replay.json

This checks every artifact hash and reloads all eight learned checkpoints. It does no training.

Limits

• Scores rank classes. They are not calibrated failure probabilities.
• Target hardware, site calibration and field acceptance are untested.
• No well group is left untouched for a future fresh test. All 40 anonymized 3W filename groups were used in this work: source qualification read all 40 (ACCESS-LEDGER.md); the models were fitted on 16 groups, selected on 8 and scored once on 8 (initial confirmation); the remaining 8, called reserved here, were never fitted, selected or scored in this repository. The later held-out-well benchmark splits the same 40 groups differently: it seals ten other groups as its test set (WELL-00001, 04, 11, 13, 15, 23, 26, 35, 37 and 41; each was a training, validation or confirmation group here) and puts the eight reserved groups in its 30-group development set.
• The input checks for malformed, stale or wrong-unit packets are not connected to these classifiers.

License

Mixed licences, hence license: other in the metadata. ATTRIBUTION.md has the sources, the changes made and the scope of each:

• authored code and trained numeric weights: MIT (LICENSE);
• tables, scalers, feature and reload probes, traces and the raw recordings in the input kits, derived from or taken from Petrobras 3W 2.0: CC BY 4.0, with attribution;
• the IBM Granite TSPulse weights and code in the preflight kit (well_health/tspulse_preflight): Apache-2.0.

The full development log, including every recovered failure, is in HISTORY.md.

Measured evaluation review — 2026-10-07

Eight retained compact checkpoints reload on 128 eligible probes each; three stateful references replay 80,534 old confirmation decisions. Later foundation/incremental strict portability failures and the eight reserved groups (never fitted, selected or scored here; see Limits for their later use elsewhere) remain separate and preserved.

Report and retained failures · Whole-profile evaluation scope. No external rank-one position or industrial deployment qualification is established. Original scientific objects and model weights are preserved.