Inheriting the wrong prior: A note for plant engineering and operations leaders evaluating industrial AI for rotary and reciprocating pumps.
Summary
Most industrial predictive maintenance was raised on inherited assumptions: smooth delivery, a stable head-flow relationship, a fleet of near-identical units, and faults that announce themselves over weeks. Positive displacement pumps break every one of those assumptions. The result is not a model that performs slightly worse. It is a model that is measuring the wrong quantity, over the wrong interval, against the wrong baseline, and reporting its confidence with a straight face.
This article sets out why that happens, what it costs, and how to tell the difference when you are evaluating machine monitoring for p.
1. The assumption nobody writes down
Every model carries a prior. In machine learning the prior is the set of assumptions baked into the architecture, the features, the sampling rate and the training data, before a single observation arrives. It is invisible in the specification sheet and decisive in the field.
The prior inside most commercial pump analytics is centrifugal. It assumes that flow varies continuously with pressure along a characteristic curve, that the fluid is a Newtonian liquid of roughly constant viscosity, that vibration energy is broadly stationary within a load state, and that health can be summarised by an overall level compared against a population.
A positive displacement pump satisfies none of this. It moves a fixed volume per revolution or per stroke, and it does so whether the downstream system wants that volume or not. Delivery is kinematic, not hydrodynamic. Pressure is an outcome imposed by the system, not a variable the pump negotiates. That single structural difference propagates into four separate failure modes for generic analytics.
| PD Pumps behaviour | Assumption | Real world behaviour |
|---|---|---|
| Delivery | Continuous, pressure-dependent | Discrete volume per cycle, pressure-independent |
| Signal character | Approximately stationary within a load state | Deterministic periodic carrier plus transients |
| Governing fluid property | Density dominant | Viscosity dominant |
| Response to a closed discharge | Recirculates, heats, survives minutes | Pressurises to destruction in fractions of a second |
| Population for learning | Fleets of identical units | Duty-specific configurations, effectively unique |
2. Failure mode one: the rhythm is read as pathology
A screw, gear, lobe, plunger or diaphragm pump delivers fluid in packets. Those packets produce a pressure ripple and a mechanical excitation at the chamber passing frequency, with a harmonic series above it and sidebands around it wherever there is timing gear meshing, valve action or asymmetry between rotors.
To a model trained on stationary rotating equipment, this is an anomaly. A strong periodic carrier with high harmonic content looks like exactly what a gearbox looks like shortly before it fails. Autoencoders trained on centrifugal data reconstruct it poorly and report large residuals. Threshold-based systems see elevated overall levels and alarm. Classifier models trained with fault taxonomies borrowed from rotating machinery map the harmonic series onto misalignment or looseness.
The failure is subtle in the other direction too. When a system is tuned to stop alarming on normal pulsation, the usual method is to widen the bands or raise the thresholds until the machine goes quiet. That desensitises the same frequency region where genuine degradation shows up first, because valve leakage, rotor clearance growth and cavitation all express themselves as changes in the shape and modulation of that carrier rather than as new frequencies elsewhere.
Worth noting who has already conceded this point. The vibration standards community did not try to stretch general rotating machinery acceptance criteria across reciprocating machines. It wrote a separate part of the standard for them, precisely because broadband velocity criteria developed for smooth rotors do not transfer to machines with reciprocating masses and valve events. The analytics industry has not yet made the equivalent move.
3. Failure mode two: viscosity is an unobserved confounder
The commercially meaningful health indicator for most positive displacement pumps is volumetric efficiency, which is set by slip: the fluid that leaks back through internal clearances instead of being delivered.
For laminar leakage through a clearance, slip flow scales roughly as
Q_slip ∝ c³ · Δp / μ
where c is the clearance, Δp the differential pressure and μ the dynamic viscosity.
Read that expression as a data scientist rather than as a hydraulics engineer and the problem is immediate. Slip is the observable. Clearance and viscosity are both unobserved. Two independent causes drive the same measurement in the same direction, and the model has access to only one equation. The system is under-determined. This is not a data volume problem. Ten more years of history will not identify a parameter that the measurement space cannot separate.
In practice this means a pump handling heavy fuel oil, molasses, polymer or crude will show a large drop in volumetric efficiency on a hot afternoon and recover it overnight, with no change whatsoever in mechanical condition. A data-driven model learns the seasonal and diurnal pattern as degradation, or learns to ignore efficiency drops altogether. Neither outcome is useful. Worse, many of these fluids are non-Newtonian, so effective viscosity depends on shear rate, which depends on speed, which the operator changes for process reasons. The confounding is not static. It moves with production.
Separating clearance growth from fluid behaviour requires a slip model with viscosity as an explicit, temperature-corrected input. That is a structural constraint imposed from physics. It is not something a network discovers from vibration spectra.
4. Failure mode three: the event is faster than the loop
Close a valve on a centrifugal pump and it recirculates, heats up, and gives you minutes. Close a valve on a positive displacement pump and it continues to displace its rated volume into a shrinking compressible space.
The pressure rise rate is governed by the effective bulk modulus of the trapped fluid and line:
dp/dt ≈ B_eff · Q / V
Take a modest case. Effective bulk modulus softened by pipe compliance and entrained gas to around 1.0 GPa, a trapped volume of 50 litres, and a delivered flow of 100 litres per minute. That gives roughly 330 bar per second. A relief valve set at 100 bar is reached from near-atmospheric in under a third of a second. Stiffer systems and smaller trapped volumes are considerably faster.
Now place a typical analytics architecture alongside that number. A wireless condition monitoring node wakes on a duty cycle and reports summarised features every five to fifteen minutes. A cloud inference service runs on a scheduled batch. A process historian stores one-second or ten-second data. The entire excursion completes between two samples. Detection latency is not a tuning parameter here; it is an architectural ceiling.
The protective function belongs in mechanical relief and in fast local logic. The point for AI is different and more interesting: if your analytics cannot resolve events at this timescale, it also cannot tell you why the relief valve lifted, how often it is lifting on partial obstruction rather than full blockage, or whether the pump is spending hours a week in a regime that is quietly consuming valve and seal life. The record of the event is the diagnostic asset, and long windows destroy it.
5. Failure mode four: there is no fleet to learn from
Data-driven maintenance works best where the same machine exists many times over. Wind turbines, motor-driven fans, process centrifugal pumps in parallel trains: these give a population, and a population gives a baseline, a transfer path and a statistically meaningful outlier.
Positive displacement pumps are configured to a duty. Rotor pitch and lead, clearance class, elastomer selection, stroke length, valve type, relief setting, drive arrangement and installed piping vary from tag to tag. Two pumps of the same model number on the same site can behave differently enough that a model trained on one produces systematic bias on the other.
This is domain shift, and it is not the mild kind. In the language of transfer learning, the source and target distributions differ in their conditional structure, not merely in their marginals. The relationship between features and health has itself changed. Fine-tuning on a handful of target samples cannot recover it, because the target site has no labelled failures to fine-tune on. Which brings us to the third layer of the problem.
6. Deep learning does not rescue this, and adds three problems of its own
Convolutional networks, recurrent architectures and autoencoders have earned their reputation on stable-state rotating machinery. Bearing fault classification from vibration imagery, sequence models on temperature and load, reconstruction-error anomaly detection on steady processes: these are real results on real assets. Applied to positive displacement machines they carry three specific liabilities.
Labels do not exist at the required density. Supervised deep learning needs many examples of each failure mode, ideally across its progression. A well-managed plant deliberately produces very few failures. The base rate for a catastrophic PD pump failure might be one event per hundred pump-years. Any model trained on that class balance will maximise accuracy by predicting health indefinitely, and will look excellent on every metric except the one that matters. Historical maintenance records rarely close the gap, since work orders record the repair, not the onset, and almost never record the operating state that produced it.
Opacity has an operating cost, not a philosophical one. An anomaly score of 0.83 tells a control room nothing actionable. It does not say which subsystem, which physical mechanism, what the margin to consequence is, or what intervention changes the trajectory. Operators respond to unexplainable alarms in a predictable way: they verify a few, find nothing, and then stop verifying. An unexplained alarm has a short half-life. Beyond the control room, opacity blocks the things reliability functions are actually accountable for, namely root cause documentation, warranty and insurance evidence, and any credible link between an alert and a change in maintenance scope.
Robustness under real conditions remains unresolved. Models trained on specific datasets and operating envelopes degrade when conditions move outside them, and industrial conditions move constantly: sensor drift, mounting changes, tank switching, batch recipe changes, ambient swings, instrument replacement. This is an open research question rather than a solved engineering practice, and treating it as solved is where most pilots quietly die.
7. A worked example: incipient cavitation in a twin screw pump
Here is the arithmetic for a cavitation episode occuring in a twin screw pump.
The setting. A twin screw pump on a heavy fuel transfer duty at 1450 rpm, drawing from a tank with a marginal static head. Fluid temperature is rising, vapour pressure is increasing, and available NPSH is eroding towards the required value. Nothing has changed on the control screen.
What actually happens first. Cavitation in a screw pump does not begin as a general rise in noise. It begins inside one angular window of the rotation, at the moment the suction chamber seals and the pressure in that chamber briefly falls below the fluid’s vapour pressure. Vapour forms, is carried along the rotor axis into rising pressure, and collapses. The collapse events are high frequency, broadband and violent, and their earliest energy sits well above 20 kHz.
The timescales.
| Quantity | Value |
|---|---|
| One revolution at 1450 rpm | 41.4 ms |
| Angular window where incipient cavitation occurs (~15°) | ~1.7 ms |
| Individual bubble collapse | tens of microseconds |
| Typical vibration analysis window | 1 s |
| Typical wireless node reporting interval | 5 to 15 min |
The dilution. For a burst of amplitude ratio r occupying duty cycle d within an averaging window, the measured overall level rises by a factor of
√(1 + d(r² − 1))
Put the numbers in. A cavitation burst five times background amplitude, occupying 1.7 ms of a one second window, raises the reported RMS by 2.0 percent. Even a burst ten times background raises it by 8.1 percent. Roll that into a ten-second average reported every fifteen minutes and a ten-fold burst moves the number by 0.8 percent.
No alarm threshold on any plant is set that tight, and no trend line survives that signal-to-noise ratio. The incident is not missed because the model is poorly trained. It is missed because averaging is a matched filter for the wrong signal. Overall level is an excellent detector for sustained change and a terrible one for low duty cycle transients, and its sensitivity degrades with the square root of duty cycle.
The band limiting compounds it. Most condition monitoring channels anti-alias and decimate to a 1 kHz or 10 kHz analysis band, because that is where centrifugal and bearing faults live. The earliest cavitation energy has been filtered out in hardware before any model sees it. The algorithm never had a chance to fail, because the evidence never arrived.
What the same event looks like under an angle-gated, physics-anchored approach. Resample the signal against shaft angle rather than time, isolate the 2 ms window in which chamber sealing occurs, and evaluate the burst content there. The same ten-fold burst now occupies most of the analysis window rather than 0.17 percent of it. The contrast rises from 8 percent to several hundred percent. Independently, compute NPSH available in real time from suction pressure, fluid temperature, vapour pressure and viscosity-corrected line losses, and compare it against a speed-corrected NPSH required. Cavitation onset becomes a residual against a calculated margin, expressed in metres of head, with a named cause and a named corrective action.
What the difference is worth. Reliability engineering calls the interval between the first detectable indication and functional failure the P-F interval. For screw pump cavitation that interval is measured in weeks, and it is spent removing metal from rotor flanks and liner bores. Material loss is cumulative and irreversible. A detection scheme that consumes eight of those weeks before its trend crosses a threshold has not detected the fault early. It has documented the damage.
8. What this means
The distinction that matters is not deep learning versus classical methods. It is whether the analytics has access to the physical variables that govern the failure, and whether it can resolve events at the timescale on which they occur.
Questions that separate the two quickly:
- What was the training population, and how many positive displacement machines were in it?
- What is the sampling rate at the sensor, the analysis bandwidth, and the end-to-end latency from event to alert?
- How does the model account for viscosity and temperature when interpreting a change in volumetric efficiency?
- Are features computed against time or against shaft angle?
- When it raises an alert, does it name a mechanism and a margin, or does it return a score?
- How was it validated on faults that were deliberately seeded and measured, rather than on faults that happened to be in someone’s history?
- What happens on the first day at a site with no failure history?
A vendor with a physics-anchored architecture answers these directly. A vendor selling a general purpose anomaly detector will move the conversation to model accuracy on a benchmark dataset. That redirection is the answer.
9. Closing
Positive displacement pumps are not harder to monitor than centrifugal pumps. They are differently constrained, and the constraints are known, written down and quantifiable. Displacement is imposed. Viscosity governs internal leakage. Consequences arrive in milliseconds. Populations are small and heterogeneous.
A model that knows none of this can still produce a number every fifteen minutes. Whether that number has anything to do with the condition of the machine is a separate question, and one worth asking before the pump answers it for you.
Notes and further reading
- On the limitations of deep learning for industrial predictive maintenance, including label scarcity, interpretability and generalisation to unseen operating conditions: Scientific Reports, https://www.nature.com/articles/s41598-025-08515-z
- ISO 20816-8 addresses vibration evaluation for reciprocating machines separately from the general rotating machinery criteria in the earlier parts of the series.
- API 674 (reciprocating positive displacement pumps) and API 676 (rotary positive displacement pumps) set out the configuration variables that make duty-specific behaviour the norm rather than the exception.
- Hydraulic Institute guidance on NPSH margin covers the distinction between NPSH required at test conditions and the margin needed in service.