Distinguishing Sensor Faults from Real Drilling Process Changes
Why reliable real-time drilling analytics must determine whether an abnormal signal reflects a bad measurement or a real change in the operation.
At a drilling rig, the difficult part is rarely seeing that a channel changed. The difficult part is deciding what the change means.
Suppose standpipe pressure unexpectedly drops. There are several possible explanations. The pressure transducer may be biased. A pump may no longer be delivering the expected output. Something in the circulating system may have changed. A rig configuration value used by an engineering model may be outdated. Or the signal may simply be capturing a short-lived operational transient.
Those possibilities have very different operational consequences.
A monitoring system that treats every unexpected measurement as a sensor problem can miss real drilling events. A system that assumes every measurement is correct can do the opposite: turn bad data into false alarms, misleading engineering calculations, and poor decisions.
The problem is therefore not simply detecting abnormal data.
It is determining whether the abnormality is in the measurement or in the physical process being measured.
This distinction has been the subject of drilling-data-quality research for more than a decade, including work presented in SPE 163726, SPE 166387, and SPE 181076.[1][2][3]
The Same Abnormal Signal Can Tell Two Different Stories
Consider a pressure measurement that suddenly disagrees with what an engineer expects.
One interpretation is straightforward: the pressure sensor is wrong.
But that conclusion assumes the underlying process has not changed.
That assumption is often reasonable in tightly controlled industrial processes. Drilling is different. The wellbore changes continuously. Hole depth increases. BHA configuration changes. Mud properties change. Pumps are serviced. The drillstring transitions among drilling, reaming, circulating, tripping and static conditions. Formation properties change with depth.
The physical relationship among measured variables can therefore change even while every sensor is operating correctly.
Ambrus, Ashok and van Oort illustrated this problem in SPE 166387 using several drilling examples.[2] One involved a managed-pressure-drilling calculation in which pump configuration changes could alter the relationship between pump strokes and actual flow. If the control logic continued using the previous pump configuration, a disagreement elsewhere in the hydraulic system could incorrectly be blamed on a pressure sensor.
Another example involved hook load during reaming. Mechanical effects such as sheave friction and changes in drag associated with drillstring rotation can produce hook-load behavior that appears counterintuitive if the operation itself is not considered.
The larger lesson extends well beyond those examples:
An unexpected measurement is not proof of a bad sensor.
It is evidence that something in the measurement-and-process system no longer behaves as expected.
That distinction matters increasingly as drilling workflows become more automated.
A human drilling engineer looking at several traces may instinctively question whether a change is operational or instrumental. Software needs a systematic way to make the same distinction.
Why Simple Alarm Limits Are Not Enough
The easiest form of data validation is a range check.
If standpipe pressure cannot reasonably be negative, a negative value is bad data. If a missing channel is represented by a null or a standard missing-value code, that can also be identified directly.
These are important checks, but they solve the easiest part of the problem.
A biased sensor can continue producing entirely believable numbers.
Imagine that a pressure sensor normally reads around 4,500 psi but gradually develops a 250-psi positive bias. Nothing about 4,750 psi is inherently impossible. A simple minimum/maximum check may never detect the problem.
Temporal methods can help remove random noise by averaging several measurements, but they have another limitation: drilling is a time-varying process. Averaging does not tell us whether the underlying process changed.
Rule-based systems encounter a similar problem. An engineer can create rules for individual situations, but the number of possible operating conditions, equipment combinations and interactions quickly becomes difficult to manage.
A more powerful form of validation comes from relational redundancy.[1][3]
The idea is simple:
Instead of asking whether one sensor looks reasonable by itself, ask whether it is reasonable relative to the other measurements and engineering relationships around it.
The Physics Provides a Second Opinion
A drilling rig contains many measurements that are physically related.
Pump strokes, pump displacement and pump efficiency are related to flow into the well.
Flow rate, fluid properties, drillstring geometry and wellbore geometry influence circulating pressure losses.[4]
Hook load is related to drillstring weight, buoyancy, trajectory, friction and operational state.
Torque is related to rotation, contact forces, drilling conditions and drillstring mechanics.
These relationships create something extremely useful for data validation: analytical redundancy.
You do not necessarily need a second physical pressure transducer to obtain another estimate of what the pressure should be.
An engineering model using other measurements can provide an independent estimate.
That estimated value will not be perfect. But it creates another piece of evidence.
Physical relationships among drilling measurements provide analytical redundancy even when duplicate sensors are unavailable.
This is one reason Bayesian networks have been studied for drilling-data validation.[1][2][3]
Despite the name, the practical idea does not require the drilling engineer to think in terms of abstract probability theory.
A Bayesian network provides a structured way to connect measurements, calculated quantities and the physical relationships between them. When new evidence arrives, confidence in the measurements and relationships can change.
The important part is not that a Bayesian network is being used.
The important part is that multiple independent pieces of drilling evidence are being evaluated together instead of judging one channel in isolation.
Sensor Health and Process Health Are Different Questions
A robust validation method should keep two questions separate:
Do we trust the sensor?
and
Do we believe the physical relationship represented by the model is still behaving normally?
That distinction is subtle but important.
Consider a simplified system containing several physically related measurements.
If one sensor begins disagreeing with several otherwise consistent relationships while the other measurements remain mutually consistent, the evidence increasingly points toward a sensor problem.
But suppose several healthy sensors change together in a physically consistent way. The sensors may all be operating correctly while the process itself has changed.
SPE 166387 describes this distinction explicitly by maintaining beliefs in both sensors and processes.[2] The approach can therefore reduce the tendency to automatically blame the instrument whenever measured behavior departs from a model.
There is also a third possibility worth recognizing:
The model context may be wrong.
The sensors may be healthy.
The physical process may be healthy.
But the model could be using the wrong pump configuration, outdated mud properties, an incorrect BHA, the wrong casing geometry or another stale contextual input.
In practice, these three situations can look surprisingly similar:
| Observation | Possible explanation |
|---|---|
| One measurement disagrees with several independent relationships | Sensor problem becomes more likely |
| Several measurements change together in a physically consistent way | Process change becomes more likely |
| Measurements are mutually consistent but disagree with the expected model | Model or contextual input should be checked |
The goal is not to force every anomaly immediately into one category.
The goal is to systematically narrow the possibilities.
Uncertainty Has to Be Part of the Calculation
Neither the sensor nor the model represents absolute truth.
A sensor has limits related to accuracy and precision.
An engineering model has uncertainty because it simplifies reality and depends on imperfect inputs.
SPE 181076 describes data validation in terms of comparing these two sources of uncertainty.[3] Instead of treating the model estimate as an exact number, the model has an expected range. The sensor reading also has an uncertainty range.
A difference between a measurement and a model prediction becomes meaningful only relative to the uncertainty in both.
That changes the question from:
"Does the sensor exactly match the model?"
to:
"Is the disagreement large enough that these two pieces of evidence are no longer reasonably consistent?"
This matters in real drilling data because noise and transients are unavoidable.
One measurement falling outside an expected range for a few seconds should not necessarily trigger the same response as a persistent disagreement lasting several minutes.
The field methodology described in SPE 181076 therefore allowed questionable behavior to accumulate before a bias fault was declared.[3]
This is a useful principle for drilling analytics in general:
Persistence matters.
So does the ability to say:
There is not enough information to determine the cause yet.
An uncertain conclusion communicated honestly is more useful than a confident but incorrect alarm.
Reliable validation combines physical consistency, supporting measurements, operational context, and persistence rather than forcing every anomaly into a binary fault decision.
Operational Context Can Make a Healthy Sensor Look Broken
Physics-based validation only works if the physics model represents the operation that is actually taking place.
That requires contextual data.
The field implementation described in SPE 181076 required information including pump specifications, mud-motor information, BHA data, casing information, mud properties and traveling-block weight in addition to the real-time sensor channels.[3]
This creates an important data-engineering problem.
Real-time sensor measurements may arrive every second.
Mud properties might be updated every few hours.
A BHA may change once during a run.
Directional surveys arrive intermittently.
Equipment specifications may come from a separate reporting system entirely.
All of these data sources have to be associated with the correct portion of the drilling operation.
Otherwise, a perfectly healthy sensor can disagree with a model simply because the model is describing the wrong physical system.
This is why synchronization and contextualization are not administrative details. They are part of the engineering calculation.
A sophisticated anomaly-detection algorithm operating on poorly synchronized or stale contextual data can be less useful than a simple plot reviewed by an experienced engineer.
Real-time drilling measurements are most useful when interpreted alongside the current operational state and well context.
A Practical Example: Standpipe Pressure Drops
Consider a hypothetical drilling interval.
The pumps are operating near a steady rate and standpipe pressure has been stable around 4,800 psi. Over several minutes, measured pressure falls toward 4,500 psi.
A simple alarm system sees one thing:
Pressure decreased by 300 psi.
A more useful engineering workflow asks a sequence of questions.
1. Is the measurement itself obviously bad?
Check for missing values, impossible values, sudden discontinuities, flat-lined readings or communication problems.
If the pressure channel suddenly jumps to zero while every other hydraulic measurement remains unchanged, the sensor or data path immediately becomes suspect.
But in our example, 4,500 psi is entirely plausible.
The diagnosis has to continue.
2. Has the system configuration changed?
Were pump liners changed?
Has pump efficiency changed?
Were mud properties recently updated?
Did the BHA or nozzle configuration change?
Is the model still using the configuration corresponding to the current operation?
If one of these inputs is stale, the expected pressure itself may be wrong.
3. Do the other measurements support a real hydraulic change?
Compare pump rate, flow-related measurements and other available hydraulic indicators.
If several measurements change together in a physically compatible direction, the evidence for a real process change increases.
If every supporting measurement behaves normally while the pressure channel alone departs from several independent expectations, the pressure measurement becomes more questionable.
4. What was the rig doing when the change occurred?
A pressure change during steady rotary drilling is different from one occurring while pumps are ramping, during a connection, during a downlink sequence or immediately after another operational transition.
The operating state changes what should be considered normal.
5. Is the disagreement transient or persistent?
A short-lived pressure transient may disappear before any action is warranted.
A persistent and growing disagreement deserves progressively more attention.
At the end of this process, the system may conclude that a sensor problem is more likely.
Or it may conclude that the measurement is probably valid and that the drilling process has changed.
Or it may conclude that there is not yet enough evidence to separate the possibilities.
All three are legitimate analytical outcomes.
An abnormal pressure trend becomes more meaningful when pump behavior, flow measurements, modeled expectations, and operational state are evaluated together.
An Alarm Should Explain Why It Exists
There is another lesson in the field work that is easy to overlook.
Even if the detection algorithm is technically good, the information still has to be trusted by the person making the decision.
In the SPE 181076 field trial, measured values and model-estimated values were displayed together when evaluating questionable channels.[3]
That is a useful human-factors principle.
Instead of displaying:
STANDPIPE PRESSURE SENSOR FAULT
show the evidence that produced the conclusion.
For example:
- measured pressure,
- expected/model pressure,
- relevant supporting channels,
- the current operational state,
- how long the disagreement has persisted,
- and whether the diagnosis is confident or uncertain.
This does not mean overwhelming the driller with every intermediate calculation.
It means exposing enough evidence that an engineer can understand why the system reached its conclusion.
Better Data Quality Is Not Just a Data-Engineering Goal
Sensor validation can sound like infrastructure work that happens before the "real" drilling analysis begins.
In reality, it affects almost everything downstream.
Mechanical specific energy depends on measured drilling parameters.
Hydraulic calculations depend on pressure, flow and contextual inputs.
Torque-and-drag models depend on hook load, torque, trajectory and drillstring information.
Automated event detection depends on patterns across multiple measurements.
Historical benchmarking depends on the assumption that measurements from different wells are sufficiently reliable to compare.
SPE 181076 notes that poor data quality creates additional work for drilling-optimization teams because historical data has to be cleaned before meaningful analysis can begin.[3]
The same principle becomes even more important as more decisions are assisted by software.
An algorithm cannot compensate for a measurement it incorrectly assumes is true.
The first level of drilling intelligence is therefore not predicting more things.
It is understanding what information can be trusted, what information is uncertain, and whether an unexpected signal belongs to the instrument or to the operation itself.
From Abnormal Data to Engineering Evidence
There are two mistakes a drilling surveillance system should avoid.
The first is trusting every sensor unconditionally.
The second is trusting the model unconditionally.
A measured value is evidence.
A model prediction is also evidence.
Neither should automatically overrule the other.
Reliable interpretation comes from combining independent measurements, physical relationships, operational context and uncertainty.
That approach does not eliminate ambiguity. In some situations, the available sensor network simply cannot uniquely distinguish a measurement problem from a process change. SPE 166387 explicitly discusses cases where the available relationships are insufficient to separate the two.[2]
Recognizing that limitation is itself a sign of a mature analytics system.
When the evidence is strong, the system should help identify what changed.
When the evidence is incomplete, it should show what is known and what remains uncertain.
And when the measurements themselves cannot be trusted, that should be established before those measurements become the basis for the next engineering decision.
References
-
Ashok, P., Ambrus, A., and van Oort, E. Automatic Sensor Data Validation: Improving the Quality and Reliability of Rig Data. SPE 163726-MS, SPE Digital Energy Conference and Exhibition, The Woodlands, Texas, 2013.
-
Ambrus, A., Ashok, P., and van Oort, E. Drilling Rig Sensor Data Validation in the Presence of Real-Time Process Variations. SPE 166387-MS, SPE Annual Technical Conference and Exhibition, New Orleans, Louisiana, 2013.
-
Ashok, P., Ambrus, A., Ramos, D., Lutteringer, J., Behounek, M., Yang, Y. L., Thetford, T., and Weaver, T. A Step by Step Approach to Improving Data Quality in Drilling Operations: Field Trials in North America. SPE 181076-MS, SPE Intelligent Energy International Conference and Exhibition, Aberdeen, United Kingdom, 2016.
-
Shahri, M., Kutlu, B., Thetford, T., et al. Adopting Physical Models in Real-Time Drilling Application: Wellbore Hydraulics. SPE 191797-MS, SPE Liquids-Rich Basins Conference—North America, Midland, Texas, 2018.