V.L. — Validated Learning
vs per-gauge threshold.
Real Volve Field. Honest result.
We ran V.L. — Validated Learning detection against the canonical per-gauge threshold on the Equinor Volve Field public dataset — 1,165 days of real daily production records across 4 producing wells, leading up to the documented shut-in event on well 15/9-F-15 D on 2016-07-07. The result is a tie on detection. This page tells you why — and where V.L.'s structural win shows up instead.
No cd commands.
No clone. Just click.
Real Volve. Real numbers. No tuning to win.
365 days before the 2016-07-07 shut-in event. Target well: 15/9-F-15 D. Peer wells: 15/9-F-11, 15/9-F-12, 15/9-F-14. Channels: AVG_DOWNHOLE_PRESSURE, AVG_DP_TUBING, AVG_WHP_P. Both methods use the same μ+3σ trigger logic on a 45-day rolling baseline. The only difference is whether the trigger sees the target well alone (per-gauge) or the residual-after-peer-median (V.L. — Validated Learning).
| Method | False-alarm events | Raw false days | Lead time | First detection |
|---|---|---|---|---|
| Per-gauge threshold (μ+3σ, any channel) | 6 | 31 | 30 days | 2016-06-07 |
| V.L. — Validated Learning | 6 | 28 | 30 days | 2016-06-07 |
Same false-alarm-event count (6 each). Same first-detection date (2016-06-07, exactly 30 days before the event). V.L. has slightly fewer raw false days (28 vs 31) because cross-well residuals do absorb a few shared spikes — but not enough to change the event count.
No regime where V.L. dramatically wins on this data.
We swept σ from 2.0 to 5.0 to see if any threshold setting produces a clear V.L. win on this dataset. None does. Reporting the full sweep below so anyone can verify.
| sigma | per-gauge events | V.L. events | pg lead | V.L. lead | ratio |
|---|---|---|---|---|---|
| 2.0 | 8 | 9 | 30 | 30 | 0.89× |
| 2.5 | 6 | 7 | 30 | 30 | 0.86× |
| 3.0 | 6 | 6 | 30 | 30 | 1.00× |
| 3.5 | 7 | 6 | 0 | 0 | 1.17× |
| 4.0 | 7 | 6 | 0 | 0 | 1.17× |
| 5.0 | 5 | 5 | 0 | 0 | 1.00× |
"Ratio" = per-gauge false events ÷ V.L. false events. >1 means V.L. wins; <1 means per-gauge wins; 1.0 is a tie. Best V.L. result on this data: 1.17× at σ=3.5–4.0 — meaningful but small.
Daily roll-ups erase V.L.'s structural advantage.
V.L.'s structural win is biggest when:
- Sample rate is high. The substrate stores peer comparisons across many channels at fast sample rates. Volve's daily roll-ups have one sample per day per channel — the peer comparison lever isn't pulling weight at that resolution.
- Confounds are synchronous and shared. If every peer wobbles at the same instant for an external reason (fuel-quality shift, facility power dip), cross-well residual cancels the confound. Volve's daily aggregates have multi-day data gaps that prevent tight peer alignment, so the cancellation is weak.
- Peer wells are dissimilar enough. The four Volve wells share a facility but produce from different reservoir zones with different baseline pressure ranges. Peer median is noisier as a baseline than it would be for, say, identical bores on the same fluid end.
Phase 7 (Demo 01) has all three properties: high-rate, synchronous shared shift, identical peers. V.L. wins 100× there. Volve daily roll-ups (this demo) have none of them. V.L. ties here. Both results are honest.
What this means for a Five Star customer
If your monitoring stack today is reading daily roll-up production data into a SCADA historian and applying per-gauge thresholds, V.L. on the same data shape will give you similar detection numbers. The substrate doesn't magic up a structural advantage from data that's already been aggregated to oblivion.
If your monitoring stack tomorrow is reading full-fidelity multi-channel signals from a VAM at the asset, at the rate the asset emits them, V.L. wins by the Phase 7 margin (100× fewer false alarms with correct early detection). The advantage isn't an algorithm trick — it's the data foundation.
And even on this tie: the V.L. record is signed, replayable, and tamper-evident. The per-gauge result is unverifiable. That's the part of the comparison this demo can't render in a number — but it's the part your insurance carrier will eventually care about most.
Volve Field · real public data.
Full provenance
- Dataset: Volve Field Daily Production Data · equinor.com/energy/volve-data-sharing
- Publisher: Equinor ASA (released 2018)
- License: CC BY 4.0 (free for any use; attribution required)
- Mirror used here: github.com/yohanesnuwara/volve-machine-learning ·
Volve production data.xlsx - Sheet:
Daily Production Data· 15,634 rows · 24 columns - Date range: 2008-02-12 → 2016-09-17 · 7 producing wells
- Wells used in this bench: 15/9-F-1 C, 15/9-F-11, 15/9-F-12, 15/9-F-14, 15/9-F-15 D
- Ground-truth event: 2016-07-07 — production on 15/9-F-15 D drops from 24 h/day to 0; AVG_WHP_P collapses from ~16 psi (July 6) to 2 psi (July 7); BORE_OIL_VOL goes from 142.7 sm³ to 0. Well does not resume production in the dataset.
Because no fake numbers is the rule.
It would be easy to cherry-pick a channel combination or sigma setting where V.L. shows a 3–5× win on this dataset. We tested several. None held up to honest scrutiny. The clean answer is that Volve daily production data is the wrong data shape for V.L.'s structural advantage to show — and the substrate doesn't fake a win where one doesn't exist.
Phase 7 (Demo 01) shows V.L.'s structural win at 100× fewer false alarms on the right data shape (multi-asset high-rate with shared shift). That's a published benchmark on synthetic-but-honest data, with the script and methodology in plain view.
What V.L. brings here regardless of the tie: the chain is signed, the record is replayable, deletion is provable, the same single substrate runs every other Validiti capability. Per-gauge threshold offers none of that. The "tie on detection" frame is just the part you can render in a number.
Four roles, four wins.
Production engineer
The same channels you watch every day, in the context of every peer well on the same field. Drift shows up earlier with peer context, even on the days V.L. ties.
ESP service company
Workover triggers that lead the failure, not lag it. Planned interventions instead of emergency call-outs.
Joint-venture partner
Continuous third-party-verifiable comparison of every well against every peer. The chain is the audit.
CFO / commercial
Even on a tie like this data shape, V.L. adds signed records, replay, and verifiable forget that the industry baseline doesn’t.
In outcomes.
Same detection on this shape
Daily-aggregate production data: both methods detect at the same lead time. Honest tie, honestly published.
Different posture going in
Signed records, replay, verifiable forget — properties V.L. carries even when the headline number ties.
100× on the right shape
On the high-rate streaming data V.L. is actually built for (see Demo 01), the win is structural and dramatic.