Predictive maintenance is the strategy that monitors the real condition of equipment in operation, through variables such as vibration, temperature, ultrasonic noise and oil contamination, to predict the failure before it happens and schedule the intervention at the right moment. Instead of replacing components by the calendar, the team follows the measured state of the asset and only intervenes when the data indicates degradation in progress.
It is the most talked-about maintenance strategy in industry, and also the one most surrounded by unrealistic expectation. This guide explains what predictive maintenance is, how it works in practice, the four most used techniques (vibration analysis, thermography, ultrasound and oil analysis) and, above all, what almost no sensor vendor tells you: the data and process prerequisites that separate a predictive program that pays for itself from a collection of ignored alarms.
What predictive maintenance is and how it works
The logic of predictive maintenance starts from a physical fact: most mechanical and electrical failures are not instantaneous. A bearing does not seize out of nowhere; it spends weeks or months degrading, and that degradation leaves measurable traces: the vibration spectrum changes, the temperature rises, ultrasonic noise appears, metallic particles show up in the lubricant.
That interval between the first detectable sign and the functional failure is what reliability engineering calls the P-F interval (from point P, the detectable potential failure, to point F, the functional failure). Everything in predictive maintenance happens inside that window: the monitoring technique detects point P, and planning uses the remaining time until point F to schedule the intervention without urgency, with the part bought within the normal lead time and the stoppage in the best production window.
The P-F window also answers a recurring operational question: how often to measure. The practical reliability rule is that the interval between measurements must be comfortably shorter than the P-F window of the monitored failure mode, so that degradation never starts and ends between two collections. Failure modes with a long window accept a monthly route; short windows push the asset toward an online sensor.
In practice, a predictive program has four steps in a cycle:
Measure: collect the condition variables of the monitored assets, by an inspection route with a portable collector or by online sensors.
Analyze: compare the measurement with the equipment's baseline and with the history, identifying degradation trends.
Decide: turn the diagnosis into a recommendation: follow up, reduce load or intervene, and within what deadline.
Execute and record: open the work order, schedule, execute and return the result to the history, closing the cycle that refines the next diagnosis.
Notice that only the first step depends on an instrument. The other three depend on process, people and organized data, and that is where predictive programs fail or pay for themselves.
It is worth registering the official name: ABRAMAN (the Brazilian Association of Maintenance and Asset Management) classifies predictive maintenance, in its National Document survey, as maintenance "by condition": carried out by tracking or controlling the equipment's parameters, in contrast to preventive maintenance "by time" and corrective maintenance "after failure or breakdown". It is the same division this guide uses.
Predictive maintenance techniques: the 4 most used in industry
Each technique sees a family of failure modes. A mature program combines two or more, because none covers everything.
Vibration analysis
It is the reference technique for rotating machines: motors, pumps, fans, gearboxes, compressors. Each mechanical defect (unbalance, misalignment, looseness, bearing defect, gear-meshing problem) produces its own signature in the frequency spectrum. The analyst compares the current spectrum with the baseline and identifies not only that something is wrong, but which component is degrading and with what severity. It can be done by periodic routes with a portable collector or by fixed sensors with continuous monitoring on the most critical assets.
Thermography
The thermographic camera reveals abnormal heating points without contact and without stopping the equipment. It is the dominant technique in electrical systems: a loose connection in a panel, an overloaded busbar, a degraded fuse and phase imbalance appear as hot spots long before they become a failure or the start of a fire. In mechanics, it reveals abnormal friction in bearings, couplings and belts; in process, fouling and failure of refractory and insulation.
Ultrasound
Industrial ultrasound captures frequencies above human hearing and converts them into an audible and measurable signal. It is the most sensitive technique for two expensive problems: leaks of compressed air and other gases (every hole in the network is money continuously going away) and incipient bearing failure, which appears in ultrasound before it appears in conventional vibration. It also detects electrical arcing and corona effect in panels and substations, and defects in steam traps.
Oil analysis
The lubricant is a sample of the inside of the machine. Periodic laboratory analysis measures three things: the health of the oil itself (viscosity, additives, oxidation), the contamination (water, dust, fuel) and the wear particles, whose composition and shape indicate which internal component is wearing out and how. It is the central technique for gearboxes, hydraulic systems, compressors and combustion engines, and it is usually the best cost-benefit entry point into predictive programs.
Other techniques complement the arsenal depending on the type of plant: electrical current analysis on motors, ultrasonic thickness measurement on piping and vessels, ferrography, borescopy and structured visual inspection. The selection criterion is always the same: which failure mode matters on this asset, and which technique detects it earliest at the lowest cost.
Types of predictive maintenance: route, online and remote monitoring
Beyond the technique, the program is differentiated by the way it collects data, and that choice defines the cost:
Route monitoring (offline): an inspector walks the assets at a defined frequency with a portable collector, camera or sample bottle. It is the entry point of almost every program: low investment, covers many assets, and the route frequency must fit comfortably within the P-F window of the monitored failure mode.
Online sensors: fixed instrumentation measuring in real time, reserved for the assets where the window between sign and failure is short or the cost of a stoppage is too high to depend on a monthly route.
Contracted remote monitoring: the analysis is done by a specialized company, which receives the data and returns the diagnosis and recommendation. It solves the lack of an internal analyst, but it does not solve the lack of process: the recommendation still needs to become an executed work order and a closed history in house.
Predictive and preventive maintenance: what is the difference
The confusion between the two is the most common conceptual mistake in the area. The difference lies in the trigger for the intervention:
| Criterion | Preventive | Predictive |
|---|---|---|
| Trigger | Time, hours or cycles (calendar/counter) | Measured condition of the equipment |
| Question it answers | How long since the last intervention? | How is this equipment right now? |
| Unnecessary intervention | Happens: replaces a part with service life to spare | Rare: only intervenes with indicated degradation |
| Random failure | Does not protect (the calendar does not see it) | Detects, if the failure mode is measurable |
| Initial investment | Low: plan and discipline | Medium to high: instrumentation, analysis and process |
| Data prerequisite | Plan and basic history | Reliable register, failure history and records |
The two do not compete: they complement each other. Preventive maintenance remains the right answer for wear that is predictable and cheap to prevent (lubrication, filters, belts), while predictive maintenance takes on the critical assets whose failure modes are measurable and whose stoppages cost dearly. The decision of which strategy to apply to each asset is the subject of our comparative guide to the types of maintenance.
Advantages of predictive maintenance (and the honest limits)
The gains of a well-implemented predictive program are concrete:
Fewer unplanned stoppages: the forming failure is detected and treated within the P-F window, before it becomes an emergency.
Fewer unnecessary interventions: components with service life to spare are not replaced by the calendar, and every avoided disassembly is also an avoided risk of a reassembly defect.
Planning with real lead time: the part bought without express freight, the stoppage scheduled in the best window, the team allocated without emergency overtime.
Extended service life: operating close to the real end of the component's life, safely, instead of discarding it on the plan's date.
Technical history that accumulates: each confirmed diagnosis refines the knowledge about that plant's failure modes.
The limits deserve the same frankness: predictive maintenance does not cover failure modes that are not measurable nor truly instantaneous failures; it requires investment in instrumentation and in people able to analyze (or a contract with someone who does); and it only works on top of a data base that most plants do not yet have. Which leads to the most important section of this guide.
Real prerequisites: what comes before the sensor
The commercial conversation about predictive maintenance starts at the sensor. The technical conversation starts three steps earlier. Before investing in monitoring, a plant needs four foundations:
1. A reliable asset register
It sounds basic and it is the most common problem: equipment without a TAG, an outdated hierarchy of functional locations, a motor moved to another position without updating the system. Monitoring condition requires knowing, without ambiguity, which asset generated each measurement. If the register is wrong, the history is born contaminated.
2. Disciplined failure recording
Predictive diagnosis learns from the history: how many times this failure mode occurred, under what conditions, with what prior symptom. If work orders close with a generic description ("fixed", "replaced bearing"), without a failure mode, without a cause and days after execution, there is no base to calibrate an alarm or to prove a gain. Reliability starts in failure recording: it is the data reported in the field, on the spot, that feeds any serious analysis.
3. Mature preventive maintenance
Predictive maintenance does not replace preventive maintenance; it leans on it. If lubrication is late, if the plan is not fulfilled and if emergencies dominate the week, the predictive program becomes just another queue of recommendations that nobody has the arms to execute. The sign of maturity is simple: planned preventive maintenance is fulfilled on time and emergency corrective maintenance is under control. Without that, the right investment is still to fix the base.
4. Defined criticality
Monitoring everything is expensive and unnecessary. The criticality matrix (impact on safety, production, quality and cost) defines which assets justify an online sensor, which enter a periodic route and which do fine with preventive and planned corrective maintenance. Without that prioritization, the program's budget gets diluted on the wrong assets.
An honest readiness test, before any sensor budget: take the last 20 closed corrective work orders and check how many have the correct equipment, the failure mode filled in, the cause identified and the real execution date. If the minority passes the test, the first investment of the predictive program is to make the data be born reliable in the field, not the measurement hardware.
When the investment pays off: the decision rule
Predictive maintenance pays off when the cost of the avoided failure exceeds the cost of detecting it. That gives a practical rule, asset by asset:
Cost of the failure: add up production downtime, emergency repair, cascade damage and safety risk. The higher it is, the more predictive maintenance pays off.
Detectability: does the dominant failure mode leave a measurable trace (vibration, heat, particle, noise)? If it does not, no sensor solves it.
P-F window: is the interval between sign and failure long enough to act? Windows of weeks favor periodic routes; short windows call for online monitoring; windows of minutes call for automatic protection, not predictive maintenance.
Frequency: a recurring failure generates a recurring return for the monitoring; a rare failure of low cost hardly pays for the program.
A hypothetical example to give scale: if the unplanned stoppage of a critical compressor costs R$ 80 thousand in lost production and emergency repair, and the monthly vibration and ultrasound route for that area costs R$ 30 thousand per year, a single avoided failure pays for more than two years of monitoring. The account changes asset by asset, and that is exactly why the criticality matrix comes before the contract.
To know whether the program is delivering, track the right maintenance KPIs: the rate of emergency corrective maintenance on the monitored assets, the accuracy of the diagnoses (failure confirmed at disassembly), the average lead time between alarm and intervention and the maintenance cost per asset. A predictive program without an indicator is an act of faith.
Examples of predictive maintenance by equipment
Critical electric motors: vibration (bearing, unbalance, misalignment) + thermography on the panel and connections + current analysis for a broken rotor bar.
Process centrifugal pumps: vibration for cavitation and bearing, ultrasound for seal and lubrication, bearing temperature trending.
Large gearboxes: oil analysis with particle counting (the first warning usually appears in the lubricant) + vibration for gear meshing.
Air compressors: oil analysis, vibration and ultrasound on the distribution network to hunt leaks, which are a continuous loss of energy.
Panels and substations: periodic thermography of connections and busbars, ultrasound for arcing and corona; contactless techniques, with the installation energized.
Overhead cranes and structures: structured inspection with photographic records and wear measurement on cables, brakes and rails, feeding a trend.
How to implement it: the realistic path in 6 steps
Fix the source data: a correct asset register and disciplined failure recording in the field. It is step zero, and it is where a mobile reporting system integrated into the ERP makes an immediate difference.
Classify the criticality: a simple matrix, applied with operations, to define where predictive maintenance enters first.
Start small and measurable: one area, two techniques (vibration and thermography, or oil analysis on the gearboxes), a periodic route before an online sensor.
Define the flow of the recommendation: who analyzes, who decides, in how long the alarm becomes a scheduled work order. An alarm with no owner is the number one failure mode of predictive programs.
Close the cycle in the history: every intervention originating from monitoring goes back to the system with the diagnosis confirmed or refuted. That is what calibrates the alarm limits and proves the gain.
Expand with the result: use the documented avoided failures to justify the next wave of assets and the eventual jump to online monitoring.
Predictive maintenance in the day-to-day of PCM
A detail that separates mature programs: the predictive recommendation enters the same planning and scheduling flow as the other work orders. The vibration alarm becomes a notification, the notification becomes a work order with a deadline compatible with the estimated P-F window, the order competes for the weekly schedule with technical priority, and the closeout records what the disassembly found. When predictive maintenance runs on the side, in the analyst's spreadsheet, the diagnosis still happens, but execution falls behind and the history does not learn. The gain of predictive maintenance materializes in the scheduling, not in the report.
Frequently Asked Questions
What is predictive maintenance?
It is the maintenance strategy that monitors the real condition of equipment in operation, using techniques such as vibration analysis, thermography, ultrasound and oil analysis, to detect forming failures and schedule the intervention before the breakdown. The trigger for the action is not the calendar, it is the measured state of the asset.
What are the most common examples of predictive maintenance?
Vibration analysis on motors, pumps and gearboxes; thermography on electrical panels and connections; ultrasound for compressed air leaks and incipient bearing failure; and oil analysis on gearboxes, compressors and hydraulic systems. In every case, the periodic measurement generates a trend, and the trend triggers the scheduled intervention.
What is the difference between predictive and preventive maintenance?
Preventive maintenance intervenes at fixed intervals of time or usage, without looking at the state of the equipment; predictive maintenance measures the real condition and only intervenes when the data indicates degradation. Preventive maintenance is simpler and covers predictable wear; predictive maintenance avoids unnecessary replacements and detects failures the calendar does not see, but requires instrumentation and reliable data.
How much does it cost to implement predictive maintenance?
It depends on the scale: periodic routes with a portable collector or a contracted analysis service cost a fraction of an online sensor system and are the usual entry point. The relevant cost is not only the instrument: it is the analyst, the decision process and the data base. The account closes when the cost of monitoring stays below the cost of the failures it avoids on the critical assets.
Before budgeting sensors, it is worth knowing what level of digital maturity your maintenance is at: identify your plant's step on the maturity ladder. And if your bottleneck is step zero, the data being born reliable in the field, book a demonstration of PM Run and see how to capture a complete failure record straight from the shop floor, integrated into SAP.
