Petrobras open-sources rare oil-well failure dataset for ML

What it does

3W is Petrobras’ first open-source project: a dataset and Python toolkit for detecting undesirable events in offshore oil wells. It packages rare, real-world well failures as Parquet files and provides standardized tooling so researchers can benchmark machine-learning approaches against actual industrial time-series data. The underlying goal is reducing the production losses and half-million-dollar-a-day maintenance bills that these events can trigger.

The interesting bit

It is unusual for a major oil company to release operational well data at all, let alone label it for public ML experimentation. The dataset draws from three distinct sources of real events, which is why the project is called 3W—three sources, one well problem.

Key highlights

• Peer-reviewed dataset of rare, real undesirable events from offshore oil wells
• Python toolkit for dataset exploration and comparative ML benchmarking
• Code under Apache 2.0 and dataset files under CC BY 4.0
• Governed by Petrobras’ Well Integrity and Flow Assurance departments
• Expanding community of researchers, startups, and other oil operators

Caveats

• Benchmarks and usage examples are still being gradually incorporated
• The toolkit contains arbitrary standardization choices made explicitly for comparative analysis
• Toolkit and dataset versions are managed independently, so compatibility requires attention

Verdict

A solid resource if you build anomaly-detection models for industrial time series or oil-and-gas monitoring. Skip it if you need a finished, drop-in detection system—this is benchmark infrastructure, not a product.