Early Warning of Stuck Pipe and Drilling Dysfunction with Label-Free Statistical Detectors

A label-free approach to a label-scarce problem

Every well is drilled once. Dysfunction events occur only a handful of times per well, and rigs, mud programmes and formations behave differently — so a model trained on one campaign’s stuck-pipe history may not transfer to the next. Instead of forcing a supervised classifier onto unsuitable data, Kwantis built statistical detectors that learn what is normal for each well from its own data and flag departures as they develop.

The result is four complementary, self-calibrating detectors covering the full well life cycle:

• On-bottom drilling dysfunction — scores torque, standpipe pressure, rotation, hookload and mechanical specific energy while making hole, identifying the channel driving each anomaly.
• Off-bottom trip / stuck-pipe risk — covers pulling pipe, running casing and working the string.
• Hydraulics — monitors the mud system whenever pumps are on, covering washout, lost circulation and an informational influx / pit-gain indicator.
• Static pit monitoring — closes the loop during flow checks, waiting-on-cement and other pumps-off periods.

Each detector recalibrates continuously against the well’s recent history — with no training set, stored event model or re-fitting when the rig, geology or mud programme changes. The approach combines robust location-scale estimation, multivariate novelty detection and sequential change-point accumulation, calibrated specifically for drilling data. Every alarm remains traceable to the physical channel that produced it.

Controlling false alarms while preserving detection sensitivity

Because a detector that cries wolf will be ignored, quiet-on-normal behaviour was treated as a primary metric rather than a byproduct of recall tuning. Pre-registered validation across a broad multi-well, multi-region cohort confirmed performance on both fronts.

• The on-bottom scorer remained silent for most drilling time, flagging only a small, controlled fraction of blocks concentrated in short, depth-localized episodes.
• On controlled synthetic dysfunction injections, it achieved complete detection once a hole section had enough history to establish a baseline, correctly naming the driving channel whenever it flagged.
• The off-bottom trip detector, validated on wells not used in development, caught most in-scope stuck-pipe episodes at a healthy margin above chance. Median warning lead was measured in hours before documentation in the daily drilling report — creating time to circulate, work the string or adjust the mud programme.
• The hydraulics detectors showed exploratory but physically consistent leads on drillstring washout and escalating lost circulation well before crew discovery. Claims remained deliberately bounded: sudden total losses and minutes-scale kicks are detected at onset, not predicted, while the influx indicator prompts crews to “verify with a flow check” rather than claiming kick detection.
• Static pit detectors addressed a documented blind spot — losses and gains during pumps-off periods — using channels already available on the well.

From statistical detection to operational action

Every flag identifies the channel or mechanism behind it, so alerts become actionable prompts — such as “look at torque and drag at this depth” — rather than unexplained red lights. Notification logic was designed with the same discipline as the statistics: interruption budgets, severity-based escalation and materiality floors were refined through repeated blind, pre-registered validation on unseen wells. Only sustained, meaningful signals reach the driller.

Outcome

By replacing an unmaintainable, label-dependent classifier with a label-free, self-calibrating statistical framework, Kwantis delivers detection that works from day one on any new well, requires no retraining as conditions change, and can give operations teams hours of actionable lead time on failure modes that often become costly non-productive time — with every alert traceable to the physical signal that produced it.