- Updated: August 24, 2026
- 2 min read
Grounded Well‑Condition Anomaly Detection on the Volve Field: Constructed Labels, a Baseline, and a Dual‑Head Model
Grounded Well‑Condition Anomaly Detection on the Volve Field
Authors: Gospel Bassey, Samuel Bassey, Vincent Fakiyesi
Real‑world oil‑field monitoring presents a stark contrast to laboratory test rigs. In production environments, sensor streams arrive without any fault log, forcing anomaly‑detection algorithms to generate their own labels. In this article we present a comprehensive study on the open Volve field dataset released by Equinor, introducing two key innovations:
- Grounded anomaly labels: Each label is cross‑checked against engineering documentation, providing a physical justification for every event.
- Dual‑head model: A compact supervised architecture that simultaneously predicts event presence and event type, extending concepts from metal‑defect detection.
Why Grounded Labels Matter
Most public benchmarks rely on artificially induced faults, which can lead to over‑optimistic performance estimates. By grounding labels in real‑world engineering constraints we ensure that the model learns patterns that are truly relevant to field operations. Our analysis shows that an unsupervised baseline, which never sees the labels, still highlights the same regions flagged by our rules, confirming the validity of the constructed labels.
Baseline and Dual‑Head Model
The unsupervised baseline uses a reconstruction‑error approach on multivariate time‑series data. The supervised dual‑head model consists of a shared encoder followed by two heads: one for binary event detection and another for multi‑class event type classification. Experiments across unseen wells demonstrate strong generalisation for event presence (F1 ≈ 0.84) and reasonable accuracy for event type (macro‑F1 ≈ 0.68). Temporal localisation is coarse but provides valuable early‑warning signals.
Results & Open Resources
All data, labels, baseline scores, trained models, and code are released under a CC‑BY‑NC‑SA 4.0 license. You can explore the dataset and reproduce the experiments on our research portal. Detailed documentation of label provenance and model assumptions is included to promote reproducibility.
Key Takeaways
- Grounded labels bridge the gap between synthetic benchmarks and real‑world field conditions.
- Unsupervised detectors can still capture meaningful anomalies without explicit supervision.
- The dual‑head architecture offers a practical balance between detection accuracy and model size for deployment on edge devices.
For further reading, visit our blog or contact the authors via the contact page.
Andrii Bidochko
CTO UBOS
Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.