Capability · Anomaly Detection
Anomaly Detection in Oil & Gas
Per-well ML models that learn each asset’s decline curve and catch deviations days before fixed SCADA thresholds. Equipment-class failure signatures for rod pump, ESP, gas-lift, and plunger lift. Steep false-alarm reduction.
The Problem: Fixed Thresholds Don't Work
Traditional SCADA systems define “normal” using static high/low thresholds set by an engineer. The same threshold applies regardless of season, well history, or operational context. This approach worked when operators had 50 wells. At scale, it creates more noise than signal.
One Threshold for Every Well
A fixed low-pressure alarm at 150 PSI makes no sense when some wells normally run at 160 and others at 400. The result: constant false positives on some wells and missed failures on others.
Blind to Seasonal and Operational Changes
After a workover, a well behaves differently for weeks. During winter, surface temperatures shift readings. Fixed thresholds cannot account for context that changes over time.
Only Catches Threshold Breaches
A well gradually declining 2% per day will never breach a fixed threshold until production has dropped significantly. By then, the damage is done and revenue is already lost.
Alarm Fatigue
When 40% of alarms are false positives, operators stop trusting the system. Real problems get buried in noise and critical issues are missed.
How WorkSync Solves It
WorkSync trains an individual ML model for every well. Each model learns that well's unique operating signature, its seasonal patterns, decline curve, response to workovers, and normal variability under different conditions.
Per-Well Baselines
Instead of a universal threshold, each well has its own learned definition of normal. A reading that is perfectly fine on Well A may be a clear anomaly on Well B.
Context-Aware Detection
Models account for time of day, ambient temperature, recent maintenance, and production trends. A pressure drop after a rod change is expected, the same drop without context is a real alarm.
Subtle Pattern Recognition
ML detects gradual degradation, cyclical anomalies, and multi-parameter correlations that no fixed threshold can capture. A slow 1% daily decline across three parameters simultaneously is invisible to SCADA but clear to the model.
Steep False-Alarm Reduction
By understanding what is truly abnormal for each specific well, the system eliminates the noise that drives alarm fatigue, letting operators focus on real problems.
Equipment-Class Failure Signatures
Different lift types fail in different ways. The model carries an equipment-class prior that biases it toward the failure signatures actually relevant to that asset. The signatures below are the ones WorkSync detects pre-failure most reliably.
Equipment class
Rod Pump
Pump fillage degradation, gas interference, fluid pound, parted rod
Detected signatures
- ▸Dynamometer card area shrinking by >15% over 7 days (fillage loss)
- ▸Card shape shifting from filled to gas-interfered (top-of-stroke signature)
- ▸Sudden bottom-load drop with maintained top-load (parted rod or pump failure)
- ▸Slow downhole-pressure drift below decline forecast band (formation issue)
Equipment class
Electric Submersible Pump (ESP)
Bearing wear, motor-temp drift, intake gas locking, downhole motor failure
Detected signatures
- ▸Motor amps rising at constant frequency (load increase, scale or bearing)
- ▸Motor temperature creeping above its historical envelope (cooling failure)
- ▸Pump intake pressure spiking with discharge falling (gas locking)
- ▸Vibration trending up before failure (industry standard early-warning signature)
Equipment class
Gas Lift
Instability, slugging, valve loading, compressor-pressure deviation
Detected signatures
- ▸Casing pressure oscillation outside expected envelope (instability or slugging)
- ▸Injection-vs-production ratio drift >20% over baseline (valve loading)
- ▸Compressor discharge pressure declining at constant rate (compressor fade)
- ▸Liquid loading signature: production drops with stable casing pressure
Equipment class
Plunger Lift
Cycle-time deviation, fall-time drift, missed arrivals, controller failure
Detected signatures
- ▸Cycle-time variance >30% over the prior week (controller-tuning issue)
- ▸Fall-time creeping up (plunger wear or wellbore drift)
- ▸Missed-arrival count rising over a 24-hour window (loading or controller)
- ▸Surface pressure ramp shape changing (valve or formation issue)
Equipment class
Surface Equipment (Compressors, Tanks, Separators)
Vibration trending, tank gauging deviation, separator level instability
Detected signatures
- ▸Compressor vibration creeping above per-unit historical band (bearing wear)
- ▸Tank gauging deviation: SCADA reading vs hand gauge spread widening
- ▸Separator level cycling outside historical envelope (dump valve or controller)
- ▸Suction/discharge pressure ratio drifting (valve issue or upstream condition)
How It Works in Practice
Scenario: An operator runs 800 wells across western Oklahoma. Their SCADA system generates 120+ alarms per day. Field crews investigate an average of 45 and find real problems on only 25.
After deploying WorkSync anomaly detection, per-well models train on 12 months of historical data. Within two weeks, the system reduces daily alerts to 30, with 28 confirmed as real issues requiring action.
Operators no longer waste half their day chasing ghosts. The alerts they do receive come with context: what changed, how it compares to the well's history, and what the economic impact is if left unaddressed.
Related
Where anomaly detection fits in the bigger discipline
The oil and gas operating model anomaly detection enables
Why static thresholds drift away from declining wells
Pricing and routing what the detection layer surfaces
Ranking the flagged exceptions
Turn anomalies into dollar-impact rankings
The intervention side of anomaly detection
Clean signal before scoring
The product that runs anomaly detection in production
Where the ML layer does its work
Frequently asked
What production engineers ask about anomaly detection.
What is anomaly detection in oil and gas?
Anomaly detection in oil and gas is the practice of using machine learning models to identify deviations from each well's expected operating behavior, production rate, pressure, runtime, temperature, vibration. Unlike fixed SCADA thresholds (which apply the same rule to every well), per-well ML models learn each asset's decline curve, seasonal patterns, and recent operational history, and flag deviations against that individualized baseline.
How is ML anomaly detection different from SCADA alarm thresholds?
Fixed SCADA thresholds use static high/low rules (e.g., "alarm if pressure < 150 PSI"). They work at small scale but fail at fleet scale because each well's normal is different, baselines drift as wells decline, and slow trends never breach the threshold. ML anomaly detection trains a model per well that learns what normal looks like for that specific asset under current operating conditions, then flags deviations as they emerge, typically days before a fixed threshold would catch them.
What kinds of failures does anomaly detection catch?
Equipment-specific signatures. Rod pumps: pump fillage degradation, gas interference, fluid pound, parted rod. Electric submersible pumps (ESPs): bearing wear, motor temperature drift, intake gas locking. Gas-lift wells: instability, slugging, compression pressure deviation. Plunger lift: cycle-time deviation, fall-time drift, missed arrivals. Surface equipment: compressor vibration trending, tank gauging deviation, separator level instability. Plus general signatures: gradual production decline below forecast, intermittent shut-ins, choke-line restriction.
How much false-alarm reduction does ML deliver versus fixed thresholds?
Field deployments at the 5,000+ well reference customer show a steep false-alarm reduction within 30 days of ML model training. Illustrative pattern: SCADA generates 120+ alarms a day with a couple dozen representing real issues; after ML deployment, the daily alert list drops to roughly the set of confirmed-real issues. The model captures the well-specific normal that fixed thresholds cannot.
Does anomaly detection require full SCADA coverage?
No. WorkSync's anomaly detection works with whatever data you have, full SCADA, partial telemetry, manual gauge readings, run tickets, lease files. Coverage gaps reduce predictive lead time but do not block deployment. The same models surface where SCADA investment would pay back fastest.
How long does it take to train per-well models?
Initial training requires 6 to 12 months of historical data per well. Most operators already have this in SCADA historians or production accounting. Models start producing usable anomalies within 1 to 2 weeks of ingest; baseline confidence stabilizes around week 4. The models continue to learn from outcomes (reinforcement learning) so accuracy compounds over time.
How does anomaly detection relate to pump-by-exception and pump-by-priority?
Anomaly detection is the engine that powers pump-by-exception (the workflow). When the ML model flags a deviation, the well joins the exception list. Pump-by-priority then ranks the exception list by economic impact, a $12,500/day deviation outranks a $90/day deviation. Anomaly detection answers "is this well deviating?" Pump-by-priority answers "in what order should we visit the deviating wells?"
Does it integrate with our existing CMMS / work-order system?
Yes. WorkSync flags anomalies and pushes them into your CMMS (Maximo, SAP PM, Oracle EAM, or others) as work orders with dollar-impact estimates, failure-mode tags, and predicted intervention windows. Your existing CMMS remains the system of record; we add the intelligence layer that turns SCADA noise into ranked work.