What is predictive maintenance, and how is it rolled out in a plant?

What is predictive maintenance?

Predictive maintenance (PdM) collects data from equipment continuously, catches the early signs of degradation and estimates remaining useful life (RUL). The goal is to win a planned intervention window before the failure occurs.

The underlying logic: rotating equipment does not fail suddenly. Bearing wear, imbalance, coupling misalignment and looseness leave a trace in the vibration and temperature signature for weeks, sometimes months. Read that trace and downtime stops being unplanned.

Predictive maintenance is not a software purchase; it is a decision chain. Unless measurement, interpretation and intervention are linked in that same chain, the system produces charts but prevents no failures.

How do corrective, preventive and predictive maintenance differ?

The three strategies answer one question differently: when is the decision to intervene made?

In practice no plant runs on a single strategy. Critical equipment is monitored, mid-importance assets run on intervals, and cheap redundant equipment is deliberately run to failure. What matters is that the split is a decision rather than an accident.

What is the P-F curve, and why does it matter?

The P-F curve is the core concept describing how a failure develops. Point P is the moment a defect first becomes detectable (potential failure). Point F is the moment the equipment can no longer perform its function (functional failure).

The time between them is the P-F interval, and the entire value of predictive maintenance sits inside it. The earlier in that interval you detect, the more time you have to plan the intervention.

The critical part: different techniques enter the curve at different points. Vibration analysis warns far earlier than a temperature rise, and heat that is noticeable by hand is usually very close to point F.

• Ultrasound / acoustic emission — the earliest indication; the breakdown of the oil film and micro-contacts are heard at this stage.
• Vibration analysis — the defect shows in the frequency signature before it grows. This is where predictive maintenance mainly operates.
• Oil analysis — once wear particles begin entering the lubricant.
• Thermography / temperature — once friction turns into heat; by now the defect has progressed.
• Audible noise — the stage the operator notices.
• Heat by touch and smoke — very close to point F; at this stage there is no time left to plan.

How does a predictive maintenance system work?

A deployment has four layers. Leave one out and the system produces data but not decisions — which is the most common way these projects fail in the field.

• Field layer — vibration and temperature sensors on the equipment; current, pressure or oil analysis where needed.
• Collection and transmission layer — the gateway that gathers sensor data. In wireless deployments, battery life, measurement interval and coverage are all decided here.
• Analysis layer — decomposing the raw vibration signal into its frequency components, comparing against thresholds and tracking the trend.
• Decision layer — an anomaly produces no value until it becomes a work order. Assigning the alarm, binding it to an SLA and closing it belongs to this layer.

Which faults does vibration analysis reveal?

Vibration is the most common way to read rotating equipment condition, because every defect leaves a trace at its own frequency. Decompose the measured signal with an FFT and the developing fault can be told apart.

Running speed is written as 1X. Most defects appear at multiples of it, or at frequencies calculated from the bearing's own geometry. Amplitude alone is not enough: the frequency is what names the defect.

How are bearing fault frequencies calculated?

A bearing's defect frequencies follow from its geometry. They are not whole multiples of running speed, which is precisely why they stand out in the spectrum and why a bearing defect is not confused with anything else.

Four inputs are needed: shaft speed (S, revolutions per second), number of rolling elements (Nb), rolling element diameter (Bd), pitch diameter (Pd) and contact angle (φ). All of these come from the bearing manufacturer's catalogue.

FTF  (cage)            = (S / 2) × [1 − (Bd / Pd) × cos φ]
BPFO (outer race)      = (Nb / 2) × S × [1 − (Bd / Pd) × cos φ]
BPFI (inner race)      = (Nb / 2) × S × [1 + (Bd / Pd) × cos φ]
BSF  (rolling element) = (Pd / 2Bd) × S × [1 − ((Bd / Pd) × cos φ)²]

What is envelope analysis, and why is it needed?

An early-stage bearing defect produces very low-energy impacts. In the same spectrum they disappear underneath high-amplitude low-frequency components such as imbalance and misalignment — a conventional FFT does not show them.

Envelope analysis solves this: the signal is first filtered through a high-frequency band, discarding the low-frequency noise. The envelope of the remaining impacts is extracted and a spectrum is taken of that envelope. The bearing defect frequencies then stand out clearly.

The practical consequence: the broadband vibration value can still look normal while the BPFO peak in the envelope spectrum has already risen. In bearing monitoring, most of the early warning comes from this method.

What are the ISO 10816 vibration limits?

ISO 10816 is the internationally accepted framework for evaluating vibration severity on rotating machines. It groups machines by power and by mounting (rigid or flexible), and places the vibration velocity measured on the machine housing (mm/s RMS) into four zones.

The zones are a decision framework rather than an absolute failure threshold. Because the limit values change with machine group and mounting, your own equipment's limits are read off the relevant table.

One point matters as much as the zone itself: movement between zones. A machine sitting steadily at the top of zone B is safer than one climbing quickly through zone B. That is why the trend is tracked alongside the threshold.

Which sensors are used, and how are they chosen?

Sensor choice follows the frequency range of the fault you want to see and the physical conditions on site. The wrong sensor still produces data — it just never shows the fault you are looking for.

• Wireless vibration and temperature sensor — quick to install, no cabling cost. Battery life and measurement interval trade against each other; suited to broad coverage of many mid-importance machines.
• Wired vibration sensor (accelerometer) — for critical equipment needing continuous, high-rate sampling. Where heat and mechanical impact are a risk, durability decides.
• Temperature — a slow, sustained rise in bearing temperature is the earliest and cheapest indicator of lubrication problems. Read together with vibration, its diagnostic power rises sharply.
• Motor current signature analysis (MCSA) — for broken rotor bars, stator faults and load imbalance. Because it installs in the panel without touching the machine, it is often the only practical option on hard-to-reach motors.
• Oil analysis — tracks wear particles, water and contamination. On gearboxes and large journal-bearing machines it shows what vibration cannot.

Where and how should a sensor be mounted?

Mounting decides as much as the sensor does. A sensor's usable frequency range depends largely on how it is fixed; the wrong mounting can push the very bearing frequencies you are looking for entirely outside the measurable range.

The general rule: mount as close to the load path as possible, on the bearing housing, with nothing in between. Mounting on a cover plate or a painted surface attenuates the signal.

How often should measurements be taken?

Measurement frequency follows the criticality of the equipment and the expected P-F interval. The basic rule: the measurement interval should be at most half the P-F interval. Otherwise a defect can appear and grow between two measurements.

On wireless sensors this rule conflicts directly with battery life. The usual answer is tiered monitoring: infrequent measurement under normal conditions, automatically tightened once a threshold is crossed.

Commissioning and the period after a major overhaul are the exception; early-failure risk is high then, so measurement is temporarily made more frequent.

Which equipment should you start with?

Instrumenting a whole plant at once is neither necessary nor economic. A criticality analysis decides what gets monitored first.

A practical way to prioritise is to answer three questions for each machine: does production stop when it stops, is there a standby unit, and how long does a replacement part take to arrive? Equipment that answers badly to all three is the first wave.

A second filter: predictive maintenance is strong on rotating equipment that degrades progressively. Do not expect the same benefit from an electronic card that fails suddenly and at random.

• Bottleneck equipment — the machines that stop the whole line.
• Equipment without redundancy — no second unit to bring online.
• Long lead-time equipment — large motors, gearboxes and compressors whose parts take weeks.
• Equipment with a repeat failure history — the ones that show up in maintenance records again and again with the same problem.
• Hard-to-reach or hazardous equipment — points where route-based measurement is impractical.

How is the return on predictive maintenance calculated?

The justification is almost always the unplanned downtime avoided. The skeleton of the calculation:

First establish the cost of one hour of unplanned downtime: the contribution margin of lost production, idle labour, restart losses, scrap, and any late-delivery penalty. Then take the unplanned downtime hours the equipment accumulated over the past year from the maintenance records.

The annual gain is those two multiplied, then multiplied by the share considered avoidable. That share should be taken conservatively: early warning catches faults that develop, not every failure.

Annual gain = hourly downtime cost × annual unplanned downtime hours × avoidable share
Payback (months) = total investment ÷ (annual gain ÷ 12)

How do you start the project?

Predictive maintenance projects fail on sequencing more often than on technology. A sequence that works:

• 1. Criticality analysis — which equipment, and on what grounds. The output is an equipment list with the reasoning written down.
• 2. Baseline measurement — record the healthy signature. Thresholds only mean something against it.
• 3. Pilot deployment — start with a limited set; verify data quality, coverage and mounting.
• 4. Thresholds and alarms — the ISO framework and the baseline are used together; neither alone is enough.
• 5. Work order integration — nothing works until it is defined who receives the alarm and within what time.
• 6. Rollout — scope widens on the strength of what the pilot measurably showed.

What are the most common mistakes?

Most failed deployments repeat the same handful of mistakes. Knowing them up front is the cheapest insurance the project has.

• Starting without a baseline — if nobody knows what normal looks like, no threshold means anything.
• Not tying alarms to work orders — alarms that pile up on a dashboard with no owner are completely ignored within weeks.
• Trying to monitor everything — as scope grows, data quality and attention fall; narrow and deep gives better results.
• Setting thresholds too tight — constant false alarms permanently destroy the team's trust in the system.
• Looking only at amplitude, never at frequency — amplitude says something changed, frequency says what changed.
• Treating mounting as a detail — expecting bearing frequencies from a magnet-mounted sensor wastes the measurement before it starts.
• Leaving the maintenance team out — they are the ones who will use it; excluded from the setup decisions, the system ends up with no owner.