A terminal had a high pressure ball valve with a condition monitoring package on it for three years. The package logged stem torque on every stroke, sent a weekly report, and generated an alarm every month or so. The valve failed in service anyway, in a way the system had recorded continuously for eleven weeks: the closing torque had drifted upward by 60 percent, the trend was monotonic, and the alarm threshold had been set at a fixed absolute value that the valve never crossed until it was already too late. The instrumentation was not the problem. The interpretation was.
Condition monitoring of valves is a field with excellent sensors and poor models. Almost any signal can be measured; very few are tied to a specific failure mechanism with a threshold that means something. What follows is an attempt to sort the measurements by what they can actually detect, to give the numbers that make a trend actionable, and to describe the small amount of modelling that turns a time series into a decision.
1. First question: is the failure observable at all?
Before choosing a sensor, it is worth listing the failure modes and asking whether each one produces a signal before it produces a leak. Some do, clearly. A stem that is stiffening from packing relaxation and corrosion produces a torque rise long before it produces a leak. A seat that is eroding produces a growing differential pressure across a closed valve. An actuator spring that is losing force produces a change in the air pressure signature at the end of the stroke.
Others are close to invisible. A pit growing through a body wall from a crevice under a gasket produces nothing measurable until it is through. A ball that has lost its surface finish in a small area produces nothing until the seat is cut. And a valve that is left in one position for years accumulates damage without ever generating a transition, which is the fundamental limitation of any technique based on stroke data.
The practical consequence is that a monitoring system should be designed around the two or three mechanisms that dominate the specific service, not around a generic list of tags. In a slurry line the dominant mechanism is seat and body wear, and the useful signals are differential pressure and torque. In a clean high pressure water line the dominant mechanism is seat damage from debris and scale, and the useful signal is again differential pressure, plus the torque at the moment of unseating. In a stem sealing application the dominant mechanism is packing relaxation, and the useful signal is a fugitive emission measurement, which no torque sensor will give you.
Failure mode | Earliest observable signal | Lead time | Sensor cost |
|---|---|---|---|
Packing relaxation and stem corrosion | Rising breakout torque, then fugitive emission | Weeks to months | Low, uses existing torque data |
Seat erosion or cutting | Differential pressure decay across a closed valve | Weeks | Moderate, two pressure transmitters |
Ball surface damage | Torque ripple within a stroke, seat leak | Weeks | Low, needs high rate torque data |
Actuator spring loss | Slower stroke, changed air pressure signature | Months | Low, two pressure taps |
Solenoid or pilot valve degradation | Longer pilot response time, intermittent strokes | Days to weeks | Very low, timing only |
Body wall loss from a crevice | None until through wall | Not detectable | Not economical |
Seal extrusion after a pressure excursion | Sudden emission increase | None, it is an event | Low, but post hoc |
2. The torque signature is the cheapest diagnostic in the plant
A ball valve stroke produces a torque curve with a shape, and the shape is a fingerprint of the mechanical condition of the valve. The curve has four regions: an initial breakout peak as the seat is unloaded and the static friction is broken, a running region where the torque reflects the seat preload, the ball and stem bearing friction, and any hydrodynamic load, a seating region at the end where the torque rises as the seat is compressed, and a final hold where the actuator stalls against its stop.
Each region maps to a mechanism. A rise in the breakout peak with a stable running region means the seat is sticking, which in a soft seated valve usually means swelling, contamination or a set taken against the ball, and in a metal seated valve means galling or thermal distortion. A rise across the whole curve means friction everywhere, which points to the stem bearing, the packing, or the grease. A ripple or a step in the running region means a local defect on the ball or the seat, and it will appear at the same angular position on every stroke, which is a strong diagnostic on its own.
The thresholds that make this actionable are relative rather than absolute, because the absolute torque varies with differential pressure and temperature. A 20 percent rise in breakout torque over a rolling baseline in comparable conditions is the point at which an investigation is warranted. A 50 percent rise, or a change in the shape of the curve rather than its level, is the point at which the valve should be planned out of service. A fixed absolute threshold, as in the opening story, is nearly useless, because the valve may spend its whole life at 30 percent of it and fail at 35.
3. If there is no torque sensor, the torque is still in the data
Most valves do not have a torque transducer, and fitting one to an existing actuator is intrusive and expensive. But the torque is present in several other signals, and all of them are easier to obtain.
The first is motor current on an electric actuator. For an induction motor operating below its breakdown point, the current is approximately proportional to the load torque, so a current transformer on one phase and a sampling rate fast enough to capture the stroke gives a torque proxy with an accuracy of a few percent after calibration. The absolute value is uncertain because of supply voltage and temperature, but the trend is excellent, and the cost is one CT and one analogue input.
The second is air pressure in a pneumatic actuator. During a stroke, the pressure on the driving side and the pressure on the exhausting side, differenced and multiplied by the piston area, gives the force the actuator is applying, and the force divided by the lever geometry gives the torque. Two pressure transducers on the cylinder ports, sampled at 100 hertz, produce a curve that is as informative as a torque trace and can be fitted to any actuator without opening it.
The third is stroke time. It is the crudest signal and it is already collected by every digital positioner. A stroke that takes 20 percent longer than its baseline under the same air supply and the same differential is a stroke with more friction, a weaker spring or a restricted exhaust, and the three can be separated by looking at the pressure signature alongside it.
Signal | What it really measures | Best sampling | Failure it finds |
|---|---|---|---|
Stem torque, direct | Total mechanical resistance | 200 Hz during stroke | Seat sticking, bearing wear, ball damage |
Motor current, electric actuator | Torque proxy, plus supply variation | 500 Hz during stroke | Same as torque, with lower absolute accuracy |
Cylinder pressure, both ports | Actuator force, seal friction | 100 Hz during stroke | Spring loss, piston seal bypass, exhaust restriction |
Stroke time and position | Integrated friction and force | 1 Hz | Anything that slows the valve |
Differential pressure | Seat tightness, flow path condition | 1 Hz, continuous | Seat leak, scale, blockage |
Acoustic emission | Leak turbulence, crack growth | 1 MHz burst capture | Through seat leak, packing leak, structural crack |
Body vibration | Flow induced excitation, cavitation | 25 kHz | Cavitation, two phase flow, support looseness |
4. Acoustic emission: powerful, and easy to misuse
A leaking seat or a leaking packing generates broadband noise from turbulence and from the collapse of small cavities in the leak path, and part of that energy is in the 100 to 400 kilohertz band where structural noise in a plant is low. A piezoelectric sensor clamped to the body can detect a leak that is far too small to see or to smell, and because the signal is continuous while the leak is continuous, a single sensor can monitor a valve with no actuation at all, which is the answer to the left-in-one-position problem.
The difficulty is attenuation and interpretation. Acoustic emission in a steel pipe attenuates at roughly 1 to 3 decibels per metre depending on frequency and on the coating, so a sensor on the body of a large valve may be far from the leak path, and a sensor on the adjacent pipe may be measuring a different valve. The signal also changes with differential pressure, temperature and the phase of the fluid, so an absolute threshold in decibels is meaningless: the only usable measure is the ratio to a baseline measured on the same valve at the same conditions.
The practical protocol is therefore comparative. Establish a baseline at commissioning with the valve closed and the normal differential applied, record the band power, and trend it. A rise of 10 to 20 decibels above baseline in the same band under the same conditions is a reason to investigate; a rise of 40 decibels is a leak. Anything else is a number that impresses a visitor and tells an engineer nothing.
5. Differential pressure is the honest seat test
A seat that leaks does not need a special test; it needs two pressure measurements. With the valve closed and pressure on one side, the pressure on the other side rises toward the upstream pressure at a rate that depends on the leak area and the downstream volume. On a 65 bar line with a downstream volume of a few litres, a seat leak of a few millilitres per minute produces a measurable rise within minutes, and the rate of rise is a direct proxy for the leak rate.
The resolution requirement is what decides whether this works. Two transmitters at 0.1 percent of span on a 100 bar range resolve 0.1 bar, which is enough to see a leak that would fail a Class VI criterion and not enough to see the very small leaks that matter for fugitive emissions. At 0.05 percent of span the resolution doubles, and the cost difference is small compared with the cost of a membrane or a product contamination event. The mistake is to use the plant's process transmitters, which are usually 0.5 or 1 percent and are often ranged to 150 percent of the operating pressure, which puts the resolution at half a bar and makes the measurement useless.
There is one trap. A closed ball valve with a liquid that can heat up will show a rising downstream pressure from thermal expansion of the trapped liquid, not from a seat leak, and a valve with a self relieving seat will show the opposite. The diagnostic only works if the temperature is stable or if the temperature is measured and the expansion term is removed.
6. The digital twin is a torque model, not a 3D picture
The phrase digital twin is used to sell software, and what it usually means in a valve context is much simpler and more useful: a physics based model of the torque, evaluated with the measured pressure, temperature and position, and compared with the measured torque. The residual, the difference between prediction and measurement, is the signal.
A workable model has three terms. The first is the seat term, proportional to the differential pressure times the projected seat area times a friction coefficient, which is a function of the seat material and the surface condition. The second is the packing and bearing term, which is a function of the gland preload and the temperature, and which is nearly constant for a given valve. The third is the thermal term, which accounts for differential expansion between the ball, the seat and the body, and which can be dominant in a valve that sees temperature cycles.
Fitted to commissioning data, such a model predicts the torque to within 10 to 15 percent across the normal operating envelope, which is enough to detect a 20 percent drift. The value is not the prediction itself but the fact that it normalises the torque for pressure and temperature, so that a change in torque at a different operating point is not mistaken for degradation. Without that normalisation, most of the variation in a torque trend is the process, not the valve.
The output is then a pair of residuals, one for breakout and one for running torque, tracked with a cumulative sum or a simple exponentially weighted mean. A cumulative sum is better than a threshold because it detects a slow drift that never exceeds any single sample limit, which is exactly the failure in the opening story.
7. Sampling, bandwidth and the data you should throw away
Most valve monitoring fails on data volume rather than on physics. A three channel acquisition at 1 kilohertz for a ten second stroke is 30,000 samples, and a valve that strokes twice a day produces 22 million samples a year, most of which describe the same healthy stroke. The useful content of a stroke is a curve with perhaps fifty features, and the right architecture stores the features and a compressed copy of the raw curve, not the raw stream.
The sampling rate has to be matched to the physics. Torque features that matter, such as the breakout peak and the seating ramp, need a few hundred hertz to be resolved properly, and 200 hertz is a reasonable compromise. Acoustic emission needs a megahertz because the signal is in the hundreds of kilohertz, but only as a burst capture on a trigger, not continuously. Vibration for cavitation detection needs 20 kilohertz or more to reach the band where cavitation noise lives. Differential pressure needs 1 hertz at most, and its value is in the trend over weeks.
The other architectural decision is where the processing happens. Sending raw waveforms from two hundred valves to a cloud platform is a bandwidth and a cost problem, and it also creates a latency problem for anything that needs to act. Edge processing that extracts the features and sends a small record is the only approach that scales, and it has the side benefit of keeping the raw data available locally for the investigation that follows an alarm.
8. From diagnostics to remaining life, carefully
The step from detecting a trend to predicting a date is where most programmes overreach. There are three families of model, and each has a domain where it works.
The first is a physical wear model, which is appropriate where the mechanism is known and the rate is driven by a measurable quantity. Seat erosion in a slurry line is the classic case: the wear rate scales with a power of the velocity, roughly between 2.5 and 3, and with the solids concentration, so a model calibrated on the first year of differential pressure data can project the remaining seat life with reasonable confidence as long as the duty does not change.
The second is a statistical degradation model, typically exponential drift fitted to the residual, with a failure threshold defined by the point at which the valve no longer performs its function. This is robust and needs no understanding of the mechanism, but it needs a threshold that is defensible, and it assumes that the degradation continues at the same rate, which is exactly what a step change in duty will break.
The third is a reliability model based on a population, a Weibull fit to time to failure across similar valves, which is useful for spares planning and useless for a single valve. The mistake is to present a population based number as a prediction for a specific asset, which is how a maintenance planner ends up with a date that has no relationship to the valve in front of him.
9. What to instrument, in order of return
If a project can afford three measurements per critical valve, they should be differential pressure, actuator air pressure and stroke time, in that order. Differential pressure finds seat leaks and blockages, which are the failures that hurt. Air pressure finds actuator and spring problems, which are the failures that stop the valve moving. Stroke time is free from the positioner and catches everything that slows down. Torque, direct or inferred, is the fourth and it is the one that separates a seat problem from a bearing problem.
Acoustic emission is worth adding where the valve is normally closed and never strokes, because it is the only technique that works without movement. Vibration is worth adding where cavitation or two phase flow is a possibility, which is a process question rather than a valve question.
The rest is discipline. A baseline recorded at commissioning, at a known pressure, temperature and differential, is worth more than any algorithm. A threshold expressed as a percentage of that baseline, with a confirmation requirement of two consecutive events to suppress noise, is worth more than a fancy model. And a named person who looks at the trend once a week is worth more than both, because the failure in the opening story was visible for eleven weeks and nobody was reading.
Requirement | Recommended value | Why |
|---|---|---|
Pressure transmitter accuracy | 0.1 percent of span, 0.05 for leak detection | 0.5 percent on a 150 percent range cannot see a seat leak |
Torque or current sampling | 200 Hz during stroke | Resolves breakout peak and seating ramp |
Air pressure sampling | 100 Hz on both cylinder ports | Separates friction from spring force |
Acoustic emission band | 100 to 400 kHz, burst capture | Above plant noise, below structural attenuation limit |
Baseline record | Three strokes at commissioning, same conditions | Every threshold is relative to it |
Alarm logic | 20 percent drift, two consecutive events | Suppresses noise, catches slow trends |
Data retention | Features forever, raw curves 12 months | Raw data is for investigation, features for trending |
10. An instrumentation clause for a valve datasheet
Where condition monitoring is specified, the valve shall be supplied with a stem torque or stem position transducer rated for the full actuator torque with a 1.5 factor, a vibration accelerometer mounted on the body within 150 millimetres of the seat pocket with a range to 10 kilohertz, two pressure connections on the actuator cylinder ports, and a differential pressure measurement across the valve using transmitters of 0.1 percent span accuracy or better. The actuator shall provide a continuous position signal and shall record stroke time. All signals shall be available on a single industrial protocol interface, and the vendor shall supply a commissioning baseline dataset comprising at least three strokes at a stated differential pressure and temperature, together with the torque model constants and the alarm thresholds derived from that baseline. Acoustic emission monitoring, where specified, shall use sensors in the 100 to 400 kilohertz band with a documented mounting method and a baseline band power record. The vendor shall state which failure modes each measurement is intended to detect and the detection threshold in engineering units.
That last sentence is the one that separates a real package from a shopping list. A supplier who cannot say what a signal is for will sell a system that measures everything and predicts nothing, and the plant will pay for the sensors, the platform and the alarms, and still lose the valve.
The J-VALVES range is built to accept that instrumentation without modification: the body has a machined pad for a vibration sensor, the stem is accessible for a torque or position transducer, and the actuator can be supplied with cylinder port taps and a position transmitter as part of the assembly. The size envelope runs from DN15 to DN1400, equivalent to 1/2 inch through 56 inch, in ASME Class 150 to 2500, with carbon steel, stainless, duplex and nickel alloy bodies for service between minus 196 and plus 650 degrees C, proof tested at 1.5 times the rating on the shell and 1.1 times on the seat. Baseline torque data at the specified differential is recorded at the works and shipped with the valve, so that the first comparison in the plant is against a number rather than against a hope, and quotations including the instrumentation scope are returned within three working hours.
11. Frequently Asked Questions
What is a valve torque signature and why does it matter?
It is the torque curve recorded over a full stroke, and it has four identifiable regions: the breakout peak as the seat unloads, the running region, the seating ramp at the end, and the stall against the stop. Each region maps to a mechanism, so a rise in breakout alone points to a sticking seat, a rise across the whole curve points to packing or bearing friction, and a ripple at a fixed angular position points to a local defect on the ball or the seat. Because the absolute torque varies with differential pressure and temperature, the useful measure is the change relative to a baseline taken at comparable conditions, with 20 percent as an investigation trigger and 50 percent as a planning trigger.
How do I monitor a valve that never moves?
Acoustic emission is the only technique that works without actuation. A leaking seat or packing generates broadband noise in the 100 to 400 kilohertz band, and a piezoelectric sensor on the body can detect a leak far below the visible threshold. The catch is that absolute decibel values mean nothing because the signal depends on differential pressure, temperature and sensor distance, with attenuation of 1 to 3 decibels per metre in steel. Establish a baseline at commissioning with the valve closed and the normal differential applied, then trend the band power; a 10 to 20 decibel rise warrants investigation and 40 decibels is a leak.
Can differential pressure really detect a seat leak?
Yes, if the transmitters are accurate enough. With the valve closed and pressure on one side, the leak fills the downstream volume and the rate of rise in downstream pressure is a direct proxy for the leak rate. Two transmitters at 0.1 percent of span on a 100 bar range resolve 0.1 bar, which is enough to see a leak that fails a Class VI criterion. Using the plant process transmitters at 0.5 or 1 percent, often ranged to 150 percent of operating pressure, puts the resolution near half a bar and makes the measurement useless. Watch for thermal expansion of trapped liquid, which mimics a leak.
What does a digital twin actually do for a ball valve?
In practice it is a physics based torque model evaluated against measured pressure, temperature and position, with the residual between prediction and measurement as the signal. A workable model has a seat term proportional to differential pressure and seat area, a nearly constant packing and bearing term, and a thermal term for differential expansion. Fitted to commissioning data it predicts within 10 to 15 percent, which is enough to detect a 20 percent drift while normalising out the process variation that otherwise dominates a torque trend. Track the residual with a cumulative sum rather than a fixed threshold, because a slow drift never exceeds any single sample limit.
How often should valve condition data be sampled?
Match the rate to the physics rather than to the platform. Torque and actuator pressure need 100 to 200 hertz during the stroke to resolve the breakout peak and the seating ramp. Acoustic emission needs a megahertz but only as a triggered burst capture. Vibration for cavitation needs 20 kilohertz or more. Differential pressure needs 1 hertz at most, because its value is in the trend over weeks. Store extracted features permanently and keep raw curves for about twelve months; a valve stroking twice a day generates tens of millions of samples a year, and almost all of them describe the same healthy stroke.
