Grounded Well-Condition Anomaly Detection on the Volve Field: Constructed Labels, a Baseline, and a Dual-Head Model

arXiv:2608.05685 · cs.AI · Submitted 2026-08-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Grounded Well-Condition Anomaly Detection on the Volve Field: Constructed Labels, a Baseline, and a Dual-Head Model".

Jane: The paper was written by Gospel Bassey, Samuel Bassey and Vincent Fakiyesi from University of Georgia.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everyone. Today we're digging into a paper that's got a mouthful of a title — "Grounded Well-Condition Anomaly Detection on the Volve Field: Constructed Labels, a Baseline, and a Dual-Head Model." Jane, I'll be honest, when I first saw "grounded well-condition anomaly detection," I thought it was about some kind of water well problem.

Jane: Ha, that's fair, Tom. But no, this is about oil and gas wells. The Volve field is an offshore oil field in the North Sea, operated by Equinor, and they released all this real production data to the public back in two thousand eighteen. It's a huge deal because real industrial data like this is usually locked away. But here's the catch — the data comes with zero labels. Nobody marked down when a well was having trouble.

Tom: Right, so the sensors recorded pressures and temperatures for years, but there's no log saying "hey, this well broke down on this date." And that's exactly the problem the authors are tackling. They had to invent the labels themselves, and then they had to be honest about the fact that they invented them.

Jane: And that's the part I love. They didn't just make up rules and pretend they were ground truth. They built rules to flag suspicious patterns, but then they checked every single candidate against the field's own engineering documents. The field development plan literally describes what can go wrong — hydrate risks during shut-ins, water breakthrough after two years, scaling that cuts productivity. So a candidate only becomes a label if the physics documented by the engineers supports it.

Tom: So it's like having a detective who finds a clue, but then has to verify it against the case file before calling it evidence. That's a much higher standard than most anomaly detection papers I've seen.

Jane: Exactly. And they released the provenance for every single label. You can look at any of the two hundred thirty-six events they flagged and see exactly why it was admitted. That kind of transparency is rare, and it's going to make this dataset a really valuable reference for anyone working on industrial condition monitoring.

Tom: And we should mention — this is the same team that did that earlier work on defect detection in metal three dee printing, right? They're clearly on a mission to make industrial AI more trustworthy.

Jane: They are. And the fact that they're building on their own prior architecture, moving it from images to time series data, that's a really interesting thread we'll get into later in the show. But for now, let's just sit with the core problem — real fields don't come with labels, and this paper is one of the more careful attempts I've seen to build them responsibly.

Tom: Stay with us, because next we're going to talk about what they actually found when they tested whether those labels were even learnable. That's where it gets really interesting.

Summary: Tom: So we're back, still on "Grounded Well-Condition Anomaly Detection on the Volve Field," and Jane, I want to get into the actual results. Because the whole paper hinges on a question — if you build labels by rule, can a model actually learn them? Or are you just chasing your own tail?

Jane: Right, and the authors were really clever about how they tested this. They ran an unsupervised baseline — an Isolation Forest — that never saw the labels at all. It just learned what "normal" looked like for each well and flagged anything unusual. And then they checked whether those flags lined up with the grounded events.

Tom: And they did. The ROC-AUC was zero point eight two five, and the precision-recall AUC was about five point eight times better than just guessing randomly. That's a big deal because it means the labels aren't arbitrary. Two completely independent methods — the rules and the unsupervised detector — agreed on which regions of the telemetry were anomalous.

Jane: And that's the convergent evidence I was talking about. It doesn't prove these are real faults, but it proves the labels correspond to actual structure in the data. The events are learnable, which is the smaller claim they're making, but it's a solid one.

Tom: But here's where I want to push back a little. They dropped some of the downhole pressure and temperature channels because those were missing about a quarter of the time. And that improved the baseline score from zero point seven seven six to zero point eight two five. But they also admit that dropping the annulus pressure channel hurt their ability to detect gas-lift instability — that event type only got caught forty-two percent of the time.

Jane: Yeah, that's a real trade-off. They gained a lot of usable data by dropping the low-coverage channels — on the longest well, usable days went from eight hundred fifty-nine to two thousand eight hundred thirty. But they paid for it in sensitivity on that one event type. I appreciate that they reported both sides of that trade instead of just celebrating the improvement.

Tom: And the gradual events — water breakthrough and productivity loss — those only have five examples each. So any result on those is basically illustrative, not conclusive. They're really upfront about that.

Jane: Which is refreshing. A lot of papers would have swept that under the rug. But here, they're saying "we have five examples, don't over-read this." That's the kind of honesty that makes me trust the rest of their numbers.

Tom: So the baseline tells us the labels are real and learnable. But the real test is the dual-head model — can it actually locate events in time and classify them? That's what we're going to dig into next.

Jane: And spoiler alert — it's a mixed bag. Some heads work great, one is just okay. We'll break down exactly what carried over from the metal-parts work and what didn't.

Improvements: Tom: Alright, we're back on "Grounded Well-Condition Anomaly Detection on the Volve Field," and we've established the labels are learnable. Now let's talk about the dual-head model — the thing that takes the architecture from their metal three dee printing work and moves it into time series. Jane, how does that work?

Jane: So imagine you're looking at a photo and you want to find a defect — you'd draw a box around it and say what it is. That's the locate-and-classify idea. Here, instead of a box in an image, they slide a sixty-day window across each well's history. The model has to say whether there's an event in that window, where it starts and ends, and what type it is.

Tom: And the results are genuinely mixed. The presence head — just "is something happening?" — that works great. On the two trustworthy folds, it scores zero point nine zero seven and zero point eight one nine. That's as good as the simpler baseline, so adding the extra heads didn't cost anything.

Jane: And the type classification is actually the surprise. It hits zero point seven nine four on the best fold, way above chance. So the model can tell a shut-in from a gas-lift instability on a well it's never seen. That's real transfer learning.

Tom: But the localization head — the one that's supposed to find where the event starts and ends — that's where it falls short. It locates events to within about seventeen days inside a sixty-day window. That's a quarter to a third of the window. Better than random, but not sharp.

Jane: And they're honest about why. Some events have fuzzy edges. A water breakthrough isn't a single crisp day — it creeps in over weeks. And the telemetry is daily and sparse. You can't expect a model to pinpoint an event to the day when the event itself doesn't have a clear boundary.

Tom: But here's what I find really interesting — they merged shut-ins and restarts into a single class. Because a restart only exists because a shut-in ended, they're one operational episode. And that's not just a modeling trick, that's physically faithful to what's happening in the well.

Jane: Right, and it recovered a bunch of data that would have been thrown away. The restart events were getting dropped because the longest-event rule kept picking the shut-in instead. Merging them was the honest fix.

Tom: So what's the improvement here? The architecture does transfer from images to time. Two of the three heads work well on unseen wells. The third one — localization — finds the events but blurs the edges. And the whole thing is only four thousand parameters on one thousand one hundred forty-one heavily overlapping windows. That's a tiny model, so the fact that it works at all is meaningful.

Jane: And the standing caveat — this recovers grounded events, not confirmed faults. Nobody from the field ever said "yes, this well was broken on this date." So it's a demonstration that the approach works, not a finished detector.

Tom: Which is exactly the right scope. They're not overselling it. But the implications for real-world deployment are huge — if you can build trustworthy labels and show a small model can learn them, that's a template for other industries with messy, unlabeled data. Next up, we're going to wrap this up and talk about what this means for the broader field.

Conclusion: Tom: We're wrapping up our discussion of "Grounded Well-Condition Anomaly Detection on the Volve Field: Constructed Labels, a Baseline, and a Dual-Head Model," and Jane, I want to pull the camera back. What's the big picture here?

Jane: The big picture is that real industrial data doesn't come with labels, and the way you build those labels matters more than the model you train. This paper shows a disciplined way to do it — rules propose, documents corroborate, and then you check your work with an unsupervised baseline. That's a template that could apply to any industry with sensor data and no fault logs.

Tom: And the dual-head model — even with its imperfect localization — shows that a tiny model can transfer across wells it's never seen. That's meaningful for real deployment, where you can't train on every possible failure mode.

Jane: Exactly. And I love that they released everything — the labels, the provenance, the code, the trained model. Anyone can audit any single label. That's how you build trust in AI systems, especially in safety-critical industries like oil and gas.

Tom: So what's the takeaway for our listeners? If you're working on anomaly detection, this is a reference benchmark worth studying. If you're working on industrial AI, it's a lesson in honesty — report your trade-offs, state your assumptions, and don't oversell your results.

Jane: And for the rest of us, it's a reminder that the hardest part of AI isn't the model — it's the data. And the people who do the unglamorous work of building trustworthy labels are the ones making real progress possible.

Tom: Well said, Jane. That's all for this paper. Thanks to everyone who tuned in, and we'll be back next time with another paper from the arXiv. Until then, keep asking questions.

Jane: And keep checking your assumptions. See you next time.

Gospel Bassey, Samuel Bassey, Vincent Fakiyesi

University of Georgia · University of Georgia

cs.AI

Submitted: 2026-08-12

Updated: 2026-08-13

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 39/100

Key concepts

Grounded Labels
These are labels created not by guessing, but by checking suspicious data patterns against the field's own engineering documents. This ensures that a flagged anomaly is supported by documented physical possibilities, providing a high standard of truth.
Unsupervised Baseline
This method, specifically using Isolation Forest, analyzed the well data to define what "normal" looked like without any pre-existing labels. It flagged anything unusual and was then compared against the grounded events to prove the labels were not arbitrary.
Dual-Head Model
This architecture adapted from image defect detection, using a sliding window across time series data. It has multiple heads designed to determine if an event is happening (presence), classify its type, and locate where it starts and ends.

Terminology

Summary

Summary

This paper addresses the problem of anomaly detection on real production field data, where fault labels do not exist and must be constructed. The authors use the open Volve field dataset released by Equinor, which contains daily production telemetry from offshore wells but no fault log. The paper makes three contributions: (1) constructing and releasing a well-condition anomaly benchmark with every label corroborated against documented physical mechanisms from the field's own development plan, including per-label provenance; (2) showing that an unsupervised detector that never sees the labels still flags the same regions the rules identified, providing convergent evidence that the labels correspond to real structure in the telemetry; (3) adapting a dual-head model from prior defect-detection work in metal additive manufacturing, extending its locate-and-classify design from spatial defects in an image to events in time.

The raw data comes from Equinor's open Volve release, specifically the production data file (Volve Production data.zip, about 2 MB) holding daily production and injection history, and the field development plan (PUD) used only as documentary evidence for labeling. The raw production data has 15,634 rows, one per well per day, from September 2007 to December 2016. The authors process this into a per-well daily table with specific decisions: zeros in pressure channels are treated as missing, not real zeros; only the six oil-producing wells are kept while injectors are set aside, leaving 9,143 daily rows; short gaps up to three days are filled by carrying the last reading forward, while longer gaps are left alone; and day-over-day changes are added alongside each channel's value.

Labels are built in two stages. Stage one uses simple rules that propose candidates: a shut-in is flagged when hours on stream drop to about zero for two or more days; a restart is the day production recovers after a shut-in ends; a water breakthrough is flagged when smoothed water cut climbs to half and stays there; a productivity loss is flagged when smoothed oil rate falls by more than thirty percent over thirty days while the choke stays open; a gas-lift instability is flagged when annulus pressure jumps far from that well's usual level. Stage two requires corroboration from the field development plan, which documents that hydrate and wax risk arises during shut-ins, gas lift is used to restart wells, injection water breaks through roughly two years into production, barium and strontium scaling and asphaltene precipitation cut productivity, and gas-lift wells show annulus-pressure instability. Only corroborated candidates become labels.

This process yields 236 grounded events: shut in (88), restart transient (79), gaslift instability (59), water breakthrough (5), and productivity loss (5). The distribution is uneven and reported as such; the two gradual events are rare, five each, roughly one per well, and any result on those is illustrative only. The authors state an honest circularity: because labels were built by rule from the same telemetry the models later read, a model trained to predict them is partly learning to reproduce those rules. The honest question is whether a model can recover the same events the grounding rules identified, which is a smaller claim than fault detection.

The paper leaves open the question of missing data. The downhole channels go absent far more often than surface channels, gaps cluster in time, and downhole pressure and temperature are always missing together while annulus pressure goes missing on different days. The authors do not treat missingness as an anomaly signal because without event-level ground truth they cannot tell whether gaps encode events or are routine gauge failure.

The baseline model is an unsupervised Isolation Forest, trained per well on normalized features. Before training, the authors measured which channels carry events and found every event type is visible through reliable surface channels, so they dropped downhole pressure, downhole temperature, and annulus pressure channels because requiring them destroyed usable days. This decision recovered a large amount of data, on the longest well taking usable days from 859 to 2,830. The same measurement showed the two gradual events barely move day-over-day change features, so the model is scoped to abrupt events. Features are standardized per well to mean zero and unit variance. The Isolation Forest prior is set to eight percent, matching the share of producer-days the grounded events cover. Five wells are trainable; well F-5 has only 144 producer-days and never has the full reliable feature set present on a single day, so it is excluded, leaving 7,692 usable well-days.

Across the five wells, with grounded events covering about seven percent of days, the unsupervised model ranks event-days above normal days with a ROC-AUC of 0.825 and a PR-AUC of 0.379, about 5.8 times better than the base rate. Dropping the low-coverage downhole channels raised the ROC-AUC from 0.776 to 0.825. Broken down by event type at an alert budget matching the eight-percent prior, the model catches shut-ins and restarts reliably (0.81 and 0.76 flagged), catches the two gradual events on three of five each, and is weakest on gas-lift instability at 0.42, a traceable cost of dropping the annulus pressure channel.

The dual-head model carries over an architecture from prior defect detection in metal additive manufacturing, moving from images to time. It slides a sixty-day window forward seven days at a time across each well, producing 1,141 windows across five wells, of which 584 contain at least one event and 557 are background. Windows more than a fifth missing are dropped; the rest are normalized per well, and gaps are filled with zero, the scaled mean. The windows overlap heavily, sharing about eighty-eight percent of their days, so the model is kept small at about four thousand parameters. Evaluation is by holding out whole wells, so no window from a test well appears in training. The model predicts one event per window, the longest one, with ties going to the earliest. Because shut-ins and restarts almost always occur together and the longest-event rule kept dropping restarts, the authors merged shut-in and restart into a single shut-in and restart cycle class, which is physically faithful and recovers the restart data. After the merge, classes are: background 557, shut-in and restart cycle 423, gas-lift instability 76, productivity loss 46, water breakthrough 39.

The architecture shares one small convolutional backbone reading the sixty-day window, then splits into three heads: one says whether an event is present, one predicts the event's start and end position in the window, and one predicts the type. The interval and type heads are masked out of losses on background windows. Results under leave-one-well-out evaluation, reading the two large-well folds F-12 and F-14 as primary evidence: presence AUC is 0.907 and 0.819, level with the simpler presence-only model; type accuracy is 0.794 and 0.511, well above chance level of about 0.20 for four classes plus background; interval localization is to within roughly seventeen days inside a sixty-day window, better than random but nowhere near sharp. The other folds give presence AUC of 0.939 (F-15D), 0.873 (F-11), and 0.881 (F-1C), with type accuracy of 0.775, 0.621, and 0.650 respectively. The authors state that presence is success, type classification is a surprise in a good way, and interval localization is where they have to be candid, noting that several event types have genuinely fuzzy edges and the telemetry is daily and sparse.

Limitations are stated plainly: one field, seven wells (six producers), daily sampling; Volve's ESP pumps are described in the field plan but never show up in the released production data; labels are constructed and corroborated but never operator-confirmed; two of five event types have only five examples each; the dual-head model runs on 1,141 windows overlapping by roughly 88 percent, so its numbers show the architecture transfers across wells and not much more; and the missing-data question stays open on purpose.

The release is public and organized as a pipeline. A single script, build dataset.py, reads the Daily Production Data sheet of Equinor's Volve production data.xlsx and produces the cleaned telemetry (volve well daily.csv and Parquet copy), the grounded events (volve grounded events.csv), and the per-label provenance file (volve events provenance.csv). The baseline notebook produces unsupervised anomaly scores (volve anomaly scores.csv), and the dual-head notebook produces the trained model (volve dualhead model.keras). Files are hosted on Hugging Face and Kaggle under CC-BY-NC-SA 4.0, matching Equinor's upstream Open Data Licence. The authors do not repost Equinor's raw files but link to Equinor's own distribution and release the code that converts the raw export into theirs. Nothing uses pretrained weights or outside training data; every value the models learn comes from the one Volve production sheet, and every label points back to a rule and a documented mechanism a reader can check against the field development plan.

The conclusion states that real fields do not come with faults marked, so labels must be built, and building labels is where quiet assumptions do their damage. The authors built anomaly labels checked against the field's own documented physics, released the reasoning behind each one, and asked whether the labels hold up. An unsupervised model that never saw them landed on the same regions, saying they are not arbitrary. A small dual-head model recovered event presence and type across wells it never saw, and located events in time only roughly, saying the architecture carries over from images to time even if the sharpest part of it does not.

Improvements for AI systems

Based on the paper, here are the specific improvements I can make to AI systems, and what the improved systems can do:

1. Grounded Label Construction with Provenance Auditing

  • Improvement: I will implement a two-stage label-construction pipeline where rules propose candidates and domain documents (e.g., field development plans, maintenance logs) corroborate them before they become training labels. Each label will carry a provenance record (well, dates, channels, type, and the exact document passage that grounds it).

  • What the improved system can do: It can generate anomaly labels on real industrial telemetry where no fault log exists, and it can defend every label against audit. This prevents the common failure where rule-generated labels carry false authority and contaminate downstream models.

2. Unsupervised Baseline as a Label-Validity Check

3. Feature Selection by Measuring Event-Signal Strength, Not Guessing

4. Per-Well Normalization Instead of Pooled Scaling

5. Dual-Head Architecture for Locate-and-Classify in Time

6. Honest Handling of Small, Overlapping Data

7. Event-Type Merging for Physically Coherent Classes

8. Explicit Handling of Missing Data as an Open Question

9. Reproducible, Auditable Release Pipeline

10. Scoping Claims to What the Data Supports

Summary of what the improved AI system can do:

  • Build auditable anomaly labels on real industrial data with no fault log.

  • Validate those labels with an unsupervised detector before supervised training.

  • Recover data by dropping non-informative, low-coverage channels.

  • Normalize per-asset to avoid cross-asset false alarms.

  • Detect, locate, and classify events on unseen assets with a small dual-head model.

  • Generalize honestly on small, overlapping data via leave-one-asset-out evaluation.

  • Merge physically coupled event types for cleaner learning.

  • Avoid spurious signals from missing data.

  • Reproduce every result from public raw data.

  • Communicate its own limitations clearly to avoid costly overconfidence.

Sources

Related papers