Predictive maintenance and condition monitoring field guide

What predictive maintenance is and why it fixes on evidence

Predictive maintenance is maintenance triggered by the measured condition of the equipment, not by a calendar and not by a breakdown. You watch the machine's own signals, the vibration of a bearing, the heat at an electrical connection, the wear metals in the oil, and you act when the data shows a failure is developing. The reading is the trigger. Calendar PM services the machine whether it needs it or not. PdM services it because the evidence says it is starting to go.

That is the whole case for it. On a calendar schedule you replace a good part because the interval came due, and you still get surprised when something fails between intervals. PdM aims at both errors at once: you stop wasting service on equipment that was fine, and you catch the bad actor before it lets go. You trade a guess for a measurement.

The work is not the sensor. It is choosing the right technique for each critical asset, baselining the healthy machine, trending the condition over time, and acting in the window between the first detectable warning and functional failure. This guide covers the ladder PdM sits on, the P-F curve that defines the window, the techniques by failure mode, and the program and records that make it pay. It builds on the preventive maintenance program guide, which covers the scheduled work PdM rides on top of, and the motor bearing guide, which goes deep on the single asset PdM catches most often.

The core idea: condition, not the calendar, not the breakdown

The one thing to hold onto is that predictive maintenance acts on the equipment's own condition, measured, instead of on a guess. A vibration trend, a thermal image, an oil sample, an ultrasonic reading: each is the machine telling you what state it is in. You let the evidence decide whether to touch it, and when. That is the difference between a reliability program and a parts-changing routine.

Two failures disappear when you do this right. The first is the wasted good part, the bearing or the belt or the motor you pulled on schedule because the interval said so, with plenty of life left in it. The second is the surprise, the asset you ran to failure that took the shaft and the production hours with it when it seized at the worst possible moment. Calendar PM solves the second poorly and causes the first. Run-to-failure causes the second. PdM is the strategy that targets both, because the trigger is the actual condition.

Hold the discipline, though. A single reading is not condition. Condition is the trend of a reading against the machine's own healthy baseline, judged against a standard like the ISO vibration severity zones and the manufacturer's limits. The evidence has to be read by someone who can read it, and the finding has to become a work order, or none of it pays. Those three, trend against a baseline, judge against a standard, turn the finding into work, are the difference between a program and a pile of data.

The maintenance ladder: reactive, preventive, predictive

There are three maintenance strategies, and a real plant runs all three on purpose, matched to the asset. Reactive maintenance is run-to-failure: you fix it when it breaks. Preventive maintenance is scheduled: you service it on a fixed calendar or runtime interval regardless of its condition that day. Predictive maintenance is condition-based: you monitor the equipment and act when the data says a fault is developing. The ladder runs from the most expensive and least planned to the most efficient and most planned.

Reactive is the most expensive way to own a critical machine, because the failure arrives unplanned, after hours, with collateral damage and a rush part. It earns its place only on cheap, redundant, non-critical equipment where running it to failure costs less than watching it. Preventive is the workhorse, predictable and schedulable, and it is the subject of the preventive maintenance program guide. Its weakness is built into the fixed interval: you over-maintain the equipment that did not need the service yet, and you under-maintain the equipment that failed early between visits.

Predictive sits at the top because it removes the guess in the interval. It does not replace preventive maintenance. Someone still changes the filter and tightens the connection. PdM tells you where to spend the hours and which interval you can safely stretch, by trending the condition on the assets where the failure is expensive enough to justify watching it. Climb the ladder where the consequence pays for the climb, and stay on the lower rungs where it does not.

What is the P-F curve?

The P-F curve is the concept that holds the whole strategy together. It maps how a failure develops over time, from the first point where the fault becomes detectable, P, the potential failure, to the point where the equipment can no longer do its job, F, the functional failure. Between those two points the machine is still running, but it is on its way out, and the condition is measurable. That stretch of time is the P-F interval, and it is the window predictive maintenance works in.

The earlier in that window you detect the fault, the more lead time you have to plan. Catch a bearing at the first rise in its high-frequency signature and you may have weeks to order the part and schedule the swap on a planned shutdown. Catch it late, by the time it is loud and hot, and you are down to days or hours and an emergency. The technique you choose sets how early you can detect, which is why the earliest-warning methods matter on the assets you most want to plan around.

The interval also drives how often you have to look. A common reliability rule is that the inspection or data-collection interval should be no longer than half the P-F interval, so you do not step over the warning between checks. P-F intervals vary enormously by failure mode: bearing wear can give weeks of warning on vibration, while an electrical arc fault can give seconds, which no route schedule will ever catch and which belongs to continuous monitoring or protection instead. Match the monitoring frequency to the interval of the failure you are watching for, not to a convenient calendar.

The condition-monitoring techniques

There is no single predictive technique. There is a toolbox, and each tool reads a different physical signal and catches a different family of faults. The skill is matching the technique to the failure mode you are worried about on a given asset. Vibration finds mechanical faults in rotating equipment. Infrared finds heat. Oil analysis finds wear and contamination inside lubricated machines. Ultrasonic finds the earliest friction, the leaks, and the electrical discharge. Motor current analysis finds electrical and rotor faults from the supply. Process and controls data finds performance drift across the whole system.

Most serious programs run more than one, because the techniques overlap and confirm each other. A bearing going bad shows in vibration, in ultrasonic, in temperature, and eventually in the oil, and two independent readings pointing the same way is a stronger call than one. The table below is the starting map: the signal, the faults it catches, and where it earns its keep. The sections after it go through each one. Treat the choices as a framework to confirm against the equipment manufacturer and a reliability engineer, because the right technique per asset depends on the machine, its duty, and its failure modes.

Vibration analysis: the workhorse for rotating equipment

Vibration analysis is the most-used predictive technique on rotating equipment, and for good reason: almost every mechanical fault on a motor, pump, fan, or compressor changes how the machine vibrates, and it does so in a way that points at the cause. A healthy machine has a vibration signature. A developing fault adds energy at frequencies set by the geometry and the speed of the part that is failing, and the analyst reads those frequencies to name the fault before it is audible.

The tool that makes this work is the FFT, the fast Fourier transform, which breaks the overall vibration into its frequency components, the spectrum. The frequency tells the fault. Imbalance shows up as a peak at one times running speed. Misalignment shows at two times, often with axial energy. Mechanical looseness throws a string of harmonics. A bearing defect rides at the specific ball-pass frequencies set by the bearing's geometry, which is how vibration finds a spalling race long before it growls. The overall vibration level is the screen; the spectrum is the diagnosis.

Bearings are the asset vibration catches most, and the motor bearing guide goes deep on why bearings cause most motor failures and how to read the failed bearing. The framework for judging whether a level is acceptable is the ISO vibration severity standard, ISO 10816 and its successor ISO 20816, which sets severity zones A through D from new-machine condition to damage occurring, measured on the non-rotating parts. Set a baseline on the healthy machine, trend against it, and let the standard and the manufacturer set the alarm levels, because every machine has its own normal.

Infrared thermography: where the heat is the fault

Infrared thermography reads surface temperature with a thermal camera, and on a lot of equipment the heat is the fault. The classic target is the electrical connection. A loose or corroded lug builds resistance, resistance builds heat, and the thermal camera shows it as a bright spot against its neighbors long before it discolors the insulation or fails. You scan the gear under load, because the fault only shows when current is flowing, and you find the loose terminal, the failing contactor, the unbalanced phase, and the overloaded conductor without opening anything live.

The electrical gear is the first target, but it is far from the only one. A bearing running hot shows on infrared, which is why a thermal check on a route catches the same thing a bearing RTD would on an instrumented machine. A steam trap that has failed open shows a temperature pattern that gives it away, and a failed-closed trap shows the opposite, so a thermographer can survey a steam system fast. Refractory hot spots, blocked condenser tubes, and overloaded motors all read as heat against a baseline.

Two cautions keep infrared honest. Emissivity and reflection fool the reading, so a shiny bus bar reads cooler than it is unless you account for it, and the comparison that matters is one phase or one connection against its identical neighbor, not the absolute number alone. And the load matters: a connection scanned at light load can look fine and still be loose, so scan under real load. Drone-mounted infrared has made roof and array and tower surveys practical at scale, which the drone and aerial-inspection work covers, but the physics of the reading is the same.

Oil analysis: the blood test of the machine

Oil analysis is the blood test of a lubricated machine. You pull a sample from a gearbox, a compressor, a hydraulic system, or a large bearing, and the lab tells you three things: what the machine is wearing, what is contaminating the oil, and whether the oil itself is still fit to lubricate. It sees inside the machine without opening it, which is exactly what you want on equipment that is expensive to take apart.

Wear metals are the first read. Spectrographic analysis measures the parts-per-million of iron, copper, chromium, lead, and the rest, and the metals point at the part that is wearing: iron from gears and races, copper from bushings and bearings, and the trend of those numbers climbing is the machine grinding itself away. Particle counting and ferrography go further on the size and shape of the debris, because a sudden crop of large particles is a different alarm than a slow rise in fine wear.

Contamination is the second read. Water destroys oil by displacing the additive and rusting the steel, dirt brings in the abrasive that starts the wear, and fuel or coolant dilution shows up as a change in viscosity. The third read is the oil's own condition: viscosity, the acid number, and the additive package, which tell you whether the lubricant has aged out and needs changing on condition instead of on a calendar. Trend the sample history, not the single result, and let the lubricant supplier and equipment manufacturer set the limits, because the alarm values depend on the oil and the machine.

Ultrasonic: the earliest warning and the leak survey

Ultrasonic monitoring listens to the high-frequency sound that friction, turbulence, and electrical discharge make, above what the ear can hear. Two jobs make it valuable. It catches a bearing fault earlier than almost anything else, because the first friction of a lubrication problem or an early defect shows up as high-frequency sound before it ever shows as vibration energy or heat. On slow-turning equipment, where vibration analysis struggles, ultrasonic is often the technique that gives the earliest warning.

The second job is the leak survey. Compressed air, gas, vacuum, and steam leaks all make broadband ultrasonic noise at the leak point, and a tech with an ultrasonic detector can walk a plant and find leaks that are bleeding money and never heard them. Compressed air leakage alone is one of the largest avoidable energy losses in a lot of plants, and an ultrasonic survey is the fast way to find it. The same tool finds failing steam traps and bad valves.

On the electrical side, ultrasonic hears the arcing, tracking, and corona inside switchgear that infrared cannot see when the fault has not yet made enough heat, so the two techniques cover different stages of the same problem. Ultrasonic is also the right way to do precision lubrication: you grease the bearing while listening, and you stop when the friction sound drops, instead of pumping by count and risking the over-greasing the motor bearing guide warns about. Like every other technique, the read takes training and a baseline to be worth anything.

Motor current signature analysis: the electrical read

Motor current signature analysis, MCSA, reads the electrical current the motor draws and finds faults in the spectrum of that current. The idea is that a fault inside the motor modulates the current it pulls, so the current carries a signature of the motor's mechanical and electrical health. The strength of the method is access: you read it at the motor control center, from the supply, without getting to the motor itself, which matters when the motor is on a roof, in a pit, or buried in a machine.

Its signature catch is the broken rotor bar. A cracked or broken bar in the rotor cage produces sideband frequencies around the line frequency in the current spectrum, and MCSA reads them, which is a fault vibration can miss. It also flags air-gap eccentricity, stator winding faults, and, with the right analysis, some bearing faults, because those too leave a mark on the current.

MCSA is a complement to vibration, not a replacement. Vibration is still the first choice for the mechanical faults on most rotating equipment, and MCSA adds the electrical and rotor faults and the convenience of reading from the panel. On a large or critical motor, the two together with infrared and oil give a fuller picture than any one alone. Confirm the diagnosis against the motor manufacturer's data, because the sideband math and the alarm thresholds depend on the motor's design.

Process data and automated fault detection

The cheapest condition data is often the data the controls already collect. Process readings, temperatures, pressures, flows, and runtimes, trended over time, tell you when performance is drifting, and the drift is a developing fault. A pump whose discharge pressure is sliding for the same speed is fouling or wearing. A heat exchanger whose approach temperature is climbing is plugging. You do not always need a new sensor; you need to trend the points you already have.

On commercial HVAC, fault detection and diagnostics, FDD, layered onto the building automation system, is the automated version of this. The analytics watch the BAS points and surface the faults that hide in plain sight: a stuck economizer, simultaneous heating and cooling, a sensor reading nonsense, equipment running off schedule, and efficiency that has drifted off its baseline. Studies put a large share of commercial HVAC energy waste down to faults like these, most of it invisible because the building still holds temperature while it wastes the energy. The preventive maintenance program guide covers the economizer and controls work these analytics point you at.

Process and FDD data is the broad, always-on, low-cost layer of a program. It is good at catching system-level drift and performance faults, and it scales because the instrumentation is already installed. It is weaker at the specific mechanical diagnosis that vibration or oil gives, so it points you at the asset and the technique that confirms the fault. Use it to aim the more specialized monitoring.

Baseline and trend: the single reading lies

The most important habit in condition monitoring is that you trend the data, you do not act on a single reading. Every machine has its own normal. A vibration level, an oil result, a temperature, that would be an alarm on one machine is healthy on another, because of how it was built, how it is mounted, and what it drives. The only way to know whether a reading is bad is to compare it to that machine's own healthy baseline.

So the first thing you do on an asset you are going to monitor is capture the baseline. Run the machine to steady temperature and load, then record the condition while it is known good: the overall vibration and the spectrum at each bearing, the temperatures, an oil sample, the ultrasonic level. That baseline is the reference every later reading is judged against. A baseline that nobody captured, or nobody can find, means every future reading is a guess about whether it is normal.

Then you alarm on the change, not the snapshot. A reading that has doubled from baseline over three months is a developing fault even if its absolute value still sits inside the ISO severity zone, and a reading that is high but flat and has been high since commissioning may just be that machine's normal. The standard and the manufacturer set the absolute limits you must not exceed; the trend against baseline is what gives you the early warning inside those limits. Watch the slope, act on the slope, and confirm the absolute value against the standard.

How do you decide which equipment gets predictive maintenance?

You cannot monitor everything, and you should not try. Predictive maintenance costs money, sensors, software, route time, and a skilled analyst, and that cost only pays back on equipment whose failure is expensive enough to justify watching it. The way you decide is criticality analysis: you rank every asset by the consequence of its failure and the likelihood it will fail, and you put the program where the rank is highest.

Consequence is the bigger driver. An asset whose failure stops production, threatens safety, or takes out something downstream is critical even if it rarely fails, because the one failure is unacceptable. Likelihood weights it further, and some criticality methods add detectability, how warning the failure gives, since a fault with no warning needs continuous monitoring or a different strategy entirely. The output is a ranked list, usually sorted into critical, important, and the rest.

Then you match the strategy to the rank, and this is where the whole ladder comes together. The critical assets get predictive maintenance, with the right technique per failure mode and the monitoring frequency set by the P-F interval. The middle gets preventive maintenance on a schedule. The cheap, redundant, non-critical equipment gets run to failure on purpose, because watching it costs more than replacing it. Spreading PdM evenly across everything is the fastest way to waste the program's budget and burn out the analyst. Rank hard, and confirm the ranking with a reliability engineer and the people who run the plant, because they know which failures actually hurt.

Reliability-centered maintenance and failure modes

Reliability-centered maintenance, RCM, is the framework that decides the right maintenance strategy for each asset by working from how it actually fails. Rather than assume every machine wants the same calendar PM, RCM asks, for each function of the asset, what are the ways it can fail, what are the consequences of each failure, and what is the most effective task to prevent or detect that specific failure mode. The answer for one failure mode might be condition monitoring, for another a scheduled replacement, and for a third run-to-failure.

The analysis tool inside RCM is failure modes and effects analysis, FMEA, sometimes extended to FMECA with a criticality rank. FMEA lists the failure modes, their effects, and their causes, which is what turns a vague worry about a machine into a specific list of what to watch for and how. A predictive technique is only useful if it detects an actual failure mode, so the FMEA is what tells you whether vibration, oil, or current is the right tool for the failure you care about.

RCM is structured and it is not free, so most teams apply the full method to their most critical systems and use a lighter version of the same thinking everywhere else. The point that carries over even without a formal study is the discipline: pick the maintenance strategy from the failure mode and its consequence, not from habit. A reliability engineer leads this work, and the standards bodies for RCM give the formal process when the asset justifies it.

The CMMS: the system of record

A predictive program lives or dies on its system of record, and that system is the CMMS, the computerized maintenance management system. It holds the asset register, the work orders, the maintenance history, the schedules, and, for a PdM program, the condition data and the trends and the alarms. Without it, the readings live in spreadsheets and tech notebooks, the trend nobody can assemble across visits is the trend nobody acts on, and the program is a collection of data points instead of a history.

The CMMS is what connects the reading to the action. A vibration alarm or an oil result that crosses a limit raises a work order, the work order carries the asset history so the planner knows what was done last time, and the closed work order writes back the as-found and as-left condition that becomes the next data point in the trend. That loop, condition in, work order out, result back in, is what makes the program improve instead of repeat.

This is the work FieldOS is built to carry for a field service team: the asset register with the baseline and the condition history on each piece of equipment, the recurring schedule that puts the route and the PM on the calendar, the work order that turns a finding into planned work with photos and readings, and the trend across visits that shows what is drifting. A program with a real record trends the equipment and proves the work. A program without one is just visits, the same point the preventive maintenance program guide makes about the service log.

Turning the finding into a planned work order

A predictive finding is worth nothing until it becomes a work order. This is the step that quietly kills more programs than any technical problem: the analyst flags the developing bearing fault, the report goes in a folder, and the bearing seizes three weeks later because nobody scheduled the repair. The deliverable of condition monitoring is not the reading. It is the planned work the reading triggers.

The value of catching the fault early is the lead time it buys. A finding in the P-F window, with weeks of warning, lets you order the right part, stage the labor, and schedule the downtime on a planned shutdown when the production loss is smallest. That is a planned work order. The same failure caught at functional failure is an emergency work order: the rush part, the overtime, the collateral damage, the production stopped at the worst time. The difference between the two is money, and the lead time is what makes it.

So the finding has to carry what the planner needs to act: the asset, the fault, how far along it is, the lead time before functional failure, and the part and labor it will take. Then it has to land in the CMMS as a work order with a priority, not as a note. Close the loop by recording what you found when you opened the machine, because that as-found condition is what calibrates the next prediction. A program that turns readings into planned work orders is a program. A program that produces reports nobody schedules is overhead.

The people: the data is worthless without the read

The instrument collects the data. A person reads it, and that read is a skill, not a button. A vibration spectrum is a graph that means nothing until someone who knows the machine and the frequencies can look at it and say that peak at two times running speed is misalignment, not imbalance, and it is getting worse. The same is true of an infrared image, an oil report, and an ultrasonic trace. Buy the sensors and skip the analyst, and you have a pile of data and no diagnosis.

The trade recognizes this with certification. ISO 18436 sets the requirements for training and certifying condition-monitoring personnel, and for vibration, ISO 18436-2 runs a four-category scheme, Category I through IV, from a technician who collects data and reads against alarms up to an analyst who can diagnose complex faults and set up a program. The category tells you what a person is qualified to call. A thermographer carries a parallel certification for infrared, and oil analysis and ultrasonic have their own. Match the certification to the work you need read.

This is also why a program needs either a trained reliability technician on staff or a service that brings one. The hardest part of building PdM is not the hardware budget. It is having someone who can turn the trend into a correct call, consistently, and who builds the baselines and the alarm levels in the first place. The data is only as good as the read, and the read is the person.

The tools: route collectors, sensors, and analyzers

The hardware spans a wide range of cost and capability, and you size it to the criticality of the asset and the budget. At the simple end are handheld meters: a vibration pen, an infrared thermometer or a thermal camera, an ultrasonic detector. They are cheap, they go anywhere, and a tech can screen a lot of equipment fast with them, which makes them the entry point for most programs.

The next step up is the route data collector, a portable analyzer that captures a full vibration spectrum and stores it against the asset in software, so a tech walking a defined route month to month builds the trend the program runs on. This is the standard tool for a route-based vibration program and the one most reliability technicians carry. Oil analysis adds the sampling kit and the lab, since the analysis itself is done off-site.

At the top are permanently installed online sensors, wired or wireless, that watch a critical machine continuously and stream the data to software that alarms on the trend. Wireless sensors have dropped in price enough that they now show up on assets that would never have justified a wired installation, which is widening what continuous monitoring covers. The choice between handheld, route, and online is the next section, and it tracks criticality more than anything else.

Online sensors vs route-based collection

There are two ways to get the data, and the right one depends on how critical and how fast-failing the asset is. Route-based collection is a tech walking a fixed route on an interval, monthly or quarterly, taking readings at each machine with a handheld or a portable collector. It is cost-effective, it covers a lot of equipment with one analyst and one tool, and it is the right choice for the large middle band of assets whose failures develop slowly enough that a monthly look catches them inside the P-F interval.

Online monitoring is always-on, with permanently installed sensors streaming data continuously to software that alarms automatically. It costs more per asset, so it is reserved for the most critical machines, the ones where downtime is expensive, access is hard, or the failure develops too fast for a monthly route to catch. Remember the P-F interval rule: if the failure can go from first warning to functional failure in less time than your route cycle, the route will step over it, and that asset needs continuous monitoring or a different strategy.

Most programs run both. Online on the critical few, route on the important many, handheld spot checks on the rest, and run-to-failure on the cheap non-critical equipment that does not earn any of it. The criticality ranking decides which asset goes in which band, which is why the ranking comes before the hardware decision, not after.

The ROI: where predictive maintenance pays

Predictive maintenance pays on the critical assets, and the return comes from a few places at once: the unplanned downtime you avoid, the secondary damage you prevent, the equipment life you extend by fixing faults early, and the swing from expensive emergency repairs to cheaper planned ones. Mature reliability programs commonly report multiples of return over the first couple of years and meaningful cuts in emergency repair and total maintenance cost, though the actual figure depends on the plant and how critical the equipment is.

It is not free, and pretending it is sets the program up to be cut. The cost is real: the sensors and software, the analyst's time and training, the route hours, and the discipline to keep the baselines and trends current. On a cheap, redundant, non-critical machine, that cost is more than the failure it prevents, which is exactly why you run those to failure and aim the PdM budget at the critical assets where the avoided downtime dwarfs the program cost.

Frame the return the way the failure frames the cost. One avoided seizure on a critical machine, with its emergency labor, its rush part, its collateral damage, and its lost production, can pay for the monitoring on that asset for years. The program's job is to find those before they happen, on the assets where they are expensive, and to leave the cheap equipment alone. That is the whole economic argument, and it is why criticality and ROI are the same conversation.

Building the program and starting small

A predictive program is built in an order, and the order matters. First, rank the assets by criticality, so you know where the program belongs. Second, for each critical asset, pick the technique that catches its failure modes, the FMEA thinking from RCM. Third, set the baseline on each one while it is healthy. Fourth, build the route or install the online sensors at a frequency set by the P-F interval. Fifth, trend the data and act on the change. Sixth, turn every finding into a work order and write the result back. That loop is the program.

Do not try to stand all of it up at once. The fastest way to kill a new program is to boil the ocean, instrument everything, drown the analyst, and produce a flood of data nobody can act on. Start small. Take your handful of most critical assets and one technique, usually vibration on the big rotating equipment, prove that it catches a real fault and pays for itself, and use that win to fund the next expansion.

Then scale deliberately. Add the next technique, the next band of assets, the next set of sensors, as the program earns the budget and the analyst's capacity grows. A program that grows from a proven core is one that the plant believes in. A program that launches huge and overwhelms everyone is one that gets cut in the first budget cycle. Start critical, start small, prove it, scale.

What to document

The record is the program. Condition monitoring is nothing but the comparison of today's reading to the machine's history, so the history has to be captured in a form you can trend. Capture the criticality ranking that put the asset in the program, the healthy baseline you set at the start, the route data or sensor stream over time, the alarms and the trends, and the work orders that the findings raised and what they found when the machine was opened.

The baseline and the trend are the two records people skip and regret. A reading with no baseline to compare against is a guess, and a string of readings nobody assembled into a trend is the early warning nobody saw. Record the asset, the technique, the measurement, the date and the load it was taken at, the limit you judged it against, and the action you took, so the next person can reproduce the call and continue the trend. This is the record FieldOS keeps against each asset over time, the same way the preventive maintenance program guide describes the service log.

Common mistakes

• Trying to monitor everything instead of ranking by criticality and targeting the critical assets.
• Acting on a single reading instead of the trend against the machine's own baseline.
• Running a program with no healthy baseline, so no reading can be judged good or bad.
• Buying the sensors and skipping the skilled analyst, leaving data that nobody can diagnose.
• Collecting data that never becomes a work order, so the fault develops to failure anyway.
• Setting the route interval longer than half the P-F interval and stepping over the warning.
• Treating predictive maintenance as a replacement for preventive maintenance instead of a layer on top of it.
• Judging a reading by its absolute value alone and ignoring the rate of change from baseline.

Field checklist

Standards and references

Predictive maintenance is governed less by a single code and more by the condition-monitoring standards, the equipment manufacturer, and the judgment of a reliability engineer. For vibration, ISO 10816 and its successor ISO 20816 set the severity zones, A through D, used to evaluate machine vibration measured on the non-rotating parts, and they are the framework most programs build their alarm and trip levels on. ISO 13373 covers vibration condition monitoring procedures, and the broader ISO 17359 gives general guidelines for condition monitoring and diagnostics of machines.

For the people, ISO 18436 sets the requirements for training and certifying condition-monitoring personnel, with Part 2 defining the four-category vibration analyst scheme and other parts covering thermography, tribology and oil analysis, and ultrasound. Treat the severity zones and any interval in this guide as a starting point to confirm against the equipment manufacturer's limits and the specific machine, because the right alarm level and the right technique depend on how that machine is built and how it fails.

The strategy side draws on the reliability literature: reliability-centered maintenance and the failure modes and effects analysis behind it, and the criticality analysis that ranks the assets. A reliability engineer leads that work on assets that justify the formal method. The points to hold onto across all of it: predictive maintenance acts on the equipment's own condition, not a guess; you trend against a baseline and act inside the P-F window; and you rank by criticality and turn the data into a work order, or the program does not pay.

Units, terms, and what they mean

Predictive maintenance carries its own vocabulary, and the same idea reads differently across a vibration report, an oil lab sheet, and a CMMS screen, so the terms are worth pinning down.

Maintenance strategy comes in three forms: reactive or run-to-failure, preventive or scheduled (PM), and predictive or condition-based (PdM). Condition monitoring is the measurement that feeds PdM. Vibration severity is read in velocity, mm/s or in/s, against the ISO zones, with the FFT spectrum giving the frequency that names the fault. Oil analysis reports wear metals in parts per million and contamination and viscosity from the same sample. Infrared reads surface temperature in degrees C or F. The P-F interval is the time from detectable warning to functional failure, and criticality is the rank of an asset by the consequence and likelihood of its failure.

Reactive / preventive / predictive

Run-to-failure; scheduled on a calendar or runtime; condition-based, triggered by measured condition

Condition monitoring

Measuring a machine's vibration, heat, oil, or sound to assess its health while it runs

P-F curve / P-F interval

The failure path from potential failure (P, first detectable) to functional failure (F); the interval between is the window to act

Vibration analysis / FFT

Reading mechanical vibration; the fast Fourier transform breaks it into a frequency spectrum that names the fault

Infrared thermography

Reading surface heat with a thermal camera to find hot connections, bearing heat, and failed steam traps

Oil analysis

The blood test of a machine: wear metals, contamination, and lubricant condition from an oil sample

Criticality / RCM

Ranking assets by consequence and likelihood of failure; reliability-centered maintenance assigns the strategy per failure mode

CMMS

Computerized maintenance management system, the system of record for assets, work orders, schedules, and condition trends

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is HVAC preventive maintenance?
• How often should you grease a motor?
• What is a data center PM program?
• What does an infrared inspection find?
• What is solar PV O&M?