gospelgit/Anomaly_Detection_with_Volve

Grounded Anomaly Detection on the Volve Field: Dataset, Labels, and Models

A well-condition anomaly-detection benchmark built on Equinor's open Volve field data, in which every anomaly label is corroborated against a documented physical mechanism from the field's own development plan. The release includes the cleaned telemetry, the grounded event labels, a per-label provenance file, baseline anomaly scores, a trained dual-head model, and the notebooks that reproduce everything.

Source and attribution

This work is derived from the Equinor open Volve dataset, released by Equinor and the former Volve licence partners (Equinor, ExxonMobil Exploration and Production Norway, Bayerngas, and the Norwegian Petroleum Directorate) under the Equinor Open Data Licence. The original data is available through Equinor's open data distribution.

This release does not redistribute Equinor's raw files. It contains only processed, derived data and the code that regenerates it from Equinor's original export. Users who want the raw source should obtain it directly from Equinor. Please attribute both Equinor (as the data owner) and this work (for the derived labels, scores, and models). The derived materials in this repository are shared under CC-BY-NC-SA 4.0, consistent with the upstream licence.

What this is, and what it is not

Volve ships real production telemetry but no fault labels. Every anomaly label here is therefore constructed: rule-derived candidates admitted only when a documented mechanism from the field development plan supports them. These are corroborated inferences, not operator-confirmed fault records. A model trained or evaluated on them tests whether telemetry alone can recover the grounded events, which is not the same as detecting confirmed faults. Treat the labels accordingly, and use the provenance file to weigh any individual label.

Files

• volve_well_daily.csv / volve_well_daily.parquet — cleaned per-well daily telemetry, 15,634 rows, seven wells, September 2007 to December 2016. Columns include hours on stream, downhole and wellhead pressures and temperatures, tubing and annulus pressure, choke opening, and oil/gas/water volumes, plus a derived water cut.
• volve_grounded_events.csv — 236 grounded anomaly events across five types (shut-in, restart transient, water breakthrough, productivity loss, gas-lift instability). Each event carries its well, onset and offset dates, the channels that triggered it, its type, and the development-plan rationale that corroborates it.
• volve_events_provenance.csv — the per-event audit trail: well, dates, channels, event type, and documented rationale, so any label can be traced and checked.
• volve_anomaly_scores.csv — per-well-day anomaly scores from an unsupervised baseline (Isolation Forest) trained without labels. Used to check that a label-blind method converges on the same regions the grounded events mark.
• CODEBOOK.md — full documentation of every column, every processing decision, and the open question about sensor missingness.
• volve_dualhead_model.keras — the trained dual-head model (event presence, event interval, event type), extending a locate-and-classify architecture from prior metal-additive-manufacturing defect-detection work to events in time. Trained on all five usable wells. Expects input shaped (N, 60, 14), prepared exactly as in the notebooks (oil producers, reduced reliable-channel feature set, per-well normalization). Raw telemetry fed in directly will not produce meaningful results.
• Notebooks — the baseline notebook and the dual-head notebook, which reproduce the full pipeline from the released CSVs: cleaning, label construction, feature analysis, training, and evaluation.

Key decisions (documented in full in the codebook and the paper)

• Zeros in pressure channels are treated as missing, not as real zeros.
• Analysis is restricted to oil-producer wells (gas-lift condition monitoring).
• Short gaps are forward-filled per well up to three days; longer gaps are left empty.
• Low-coverage downhole channels are dropped from the model after a measured feature-dependence analysis showed no event type depends on them alone.
• Features are normalized within each well.
• Labels are constructed and corroborated, and the paper and codebook state every rule threshold as a declared assumption.

Results at a glance

• The unsupervised baseline, trained without labels, ranks grounded-event days above normal days with ROC-AUC 0.825 (about six times the base-rate precision). Because it never saw the labels, this convergence is evidence the grounded events are real structure, not rule artifacts.
• The dual-head model transfers across held-out wells: event presence and event type recover well, while interval localization is learnable but coarse (about a seventeen-day mean error in a sixty-day window). It is a demonstration that the architecture carries over, not a state-of-the-art detector.

Citation

If you use this release, please cite the accompanying paper and attribute Equinor as the source of the underlying field data. (Paper citation to be added on publication.)