labeled well events

A labeled well event is a stretch of sensor time series - pressures, temperatures, flow readings - whose rows carry a class label naming the abnormal condition occurring in that window. Labels turn raw telemetry into supervised training material: a model learns what developing problems look like instead of merely flagging deviation from normal.

Petrobras' 3W Dataset 2.0.0 (Oil Well Events) documents the construction pattern. It targets 'rare undesirable events in oil wells - real, simulated and hand-drawn instances from 42 real wells across 10 event classes'. The three origins matter: real instances carry field noise, simulated instances cover situations too dangerous or rare to wait for, and hand-drawn instances let engineers sketch boundary cases directly.

Distribution is file-per-instance: about 2,228 Parquet instance files organized across 10 event directories, delivered as a 1.79 GB compressed single ZIP download, with tags 'labeled events', 'anomaly detection' and 'well control'.

Anomaly detection without labels can only tell you something unusual happened; labeled events tell you what. For well control, that difference decides whether an alarm says 'investigate' or names the failure mode already underway.

Labels also encode rarity honestly. Undesirable well events are rare by nature, so a dataset that mixes real, simulated and hand-drawn instances - as 3W does across its 10 classes - lets you train against classes the field rarely produces. Buyers should still ask which share of each class is synthetic: a model validated only on simulated windows may meet different physics downhole.

Finally, labeling quality caps everything downstream. Per-instance Parquet files across 10 event directories make class balance auditable file-by-file, which is exactly what you want before committing a monitoring system to act on the model's outputs.

1. Count the classes and instances per directory. 3W spreads roughly 2,228 Parquet instance files across 10 event directories - walk the tree and confirm each class you care about has enough windows.

1. Ask how many wells contributed. The 42 real wells behind 3W set the diversity ceiling; fewer wells means less variation in equipment and reservoir behaviour.

1. Separate real from simulated from hand-drawn. The provenance mix determines how much field noise your model sees during training.

1. Check the file format suits your stack. Parquet instance files read columnar and fast; the 1.79 GB compressed ZIP download makes the whole set practical to mirror.

1. Tie labels to sensor context. Water-cut and pressure trajectories around each window should justify its label; unexplained labels cannot be audited.

Entries connected to labeled well events:

• pre-salt - the Brazilian geological setting behind Petrobras' well fleet.
• water-cut - one of the sensor signals whose drift accompanies labeled events.