Predictive maintenance (PdM) is a strategy that monitors the actual condition of equipment, vibration, temperature, oil chemistry, electrical current, and uses trends in that data to forecast when failure will occur, so the repair is scheduled just before it, with parts staged and production warned. Fix it because the data says it is failing, not because the calendar says it is due.

PdM is the most oversold concept in maintenance, so this guide is deliberately concrete: the four core technologies and what each one actually detects, the P-F curve that explains why early detection buys planning time, the prerequisites vendors tend to skip, and the savings numbers that come from a primary source rather than a brochure.

How is predictive maintenance different from preventive?

Preventive maintenance services equipment on a fixed trigger, time or usage, regardless of condition; predictive maintenance services equipment when measured condition says degradation has started. A PM schedule replaces a bearing every 12 months whether it needs it or not; PdM watches the bearing's vibration signature and replaces it in month 17, three weeks before the trend line says it would have seized.

The middle rung between them is condition-based maintenance: act when a threshold is crossed, without necessarily forecasting a failure date. PdM adds the forecast, trend analysis, and increasingly machine-learning models, on top of condition data. Where each strategy belongs per asset is mapped in our equipment reliability maturity ladder.

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

The P-F curve plots an asset's health over time between two points: P the moment a developing failure first becomes detectable, and F functional failure, the machine can no longer do its job. The gap between them, the P-F interval, is your warning time. The entire economic case for PdM is that different technologies detect failure at different points on that curve, and earlier detection means more time to plan, order parts, and schedule the repair into planned downtime.

The P-F curve: detection windows by technologyThe P-F curveasset conditiontime →P · failure detectableF · functional failure vibration analysis oil analysis thermal / infrared audible noise hot to touchP-F interval = your planning window
The P-F curve. Vibration analysis typically detects developing failures earliest; by the time a bearing is audibly noisy or hot to the touch, most of the planning window is gone. Positions are illustrative, actual intervals vary by failure mode.

The practical takeaway: PdM does not prevent the failure mode. It buys you weeks of warning instead of minutes, which converts an emergency (overtime, air-freighted parts, collateral damage, unplanned line stop) into a scheduled job.

What are the four core PdM technologies?

Four technologies cover the large majority of industrial PdM programs. Each sees a different family of failure modes.

TechnologyWhat it detectsTypical assetsNotes on cost and skill
Vibration analysisBearing wear, imbalance, misalignment, looseness, gear defectsRotating equipment: motors, pumps, fans, gearboxes, compressorsThe workhorse of PdM. Route-based readings with a handheld analyzer, or permanent wireless sensors on critical assets. Interpretation takes trained analysts (certification programs exist) or increasingly software-assisted analysis.
Thermal / infrared imagingHot spots from loose electrical connections, overloaded circuits, failing insulation, friction, refractory lossElectrical panels, switchgear, motor control centers, steam systems, furnacesFast to survey, relatively low entry cost for a camera and training. Many failure modes show heat late on the P-F curve, so pair it with earlier-detection methods on critical rotating assets.
Oil analysisWear-metal particles, contamination (water, dirt), lubricant degradation, incorrect lubricantGearboxes, hydraulics, engines, compressors, large lubricated bearingsLab-based sampling on an interval, typically modest cost per sample. Doubles as a check on your lubrication program many findings are contamination problems, not machine problems.
Current / motor circuit analysisRotor bar defects, stator winding degradation, voltage imbalance, load anomaliesElectric motors and driven equipmentCan piggyback on electrical data already flowing from drives and PLCs. Attractive where sensors are hard to mount; needs electrical expertise to interpret.
PdM technology mapFour technologies, one assetCRITICAL ASSETmotor · gearbox · pumpVIBRATIONbearings · imbalance · alignmentTHERMAL / IRhot spots · connections · frictionOIL ANALYSISwear metals · contaminationCURRENT ANALYSISrotor · windings · load anomaliesreadings → baseline → trend → forecast → planned work order
The four core condition-monitoring technologies. Most programs start with vibration on rotating equipment plus thermal surveys on electrical gear, then add oil and current analysis where the asset base justifies them.

What does PdM actually require? The honest prerequisites

PdM fails more often from missing foundations than from bad sensors. Before buying anything, work through this readiness sequence.

  1. An asset register with criticality rankings. PdM is applied per asset, and only to assets whose failure consequence justifies the monitoring cost. If you cannot rank your assets, you cannot pick monitoring candidates.
  2. Failure history worth trusting. You need to know which failure modes actually hurt you, from work-order history and downtime tracking with reason codes. Monitoring for failure modes you do not have is a common and expensive mistake.
  3. A functioning PM program. If the crew cannot hold 90% PM compliance it will not act on PdM alerts either. PdM generates work; a plant drowning in reactive work has no capacity to receive it.
  4. Baselines and data plumbing. A reading means nothing without a healthy-state baseline and a trend. That means repeatable measurement points, consistent collection (routes or permanent sensors), and data that lands somewhere it can be trended, not a folder of PDFs. Plants that already stream machine data from PLCs and sensors into one operational layer, the approach Harmony takes with no rip-and-replace, have most of this plumbing in place.
  5. A named owner who closes the loop. Someone must own the readings, screen the alerts, and convert real ones into planned work orders. An alert that does not become a work order within the P-F window was just an expensive notification. This is where planning and scheduling discipline pays off.
  6. A pilot scope, not a plant-wide rollout. Pick 10–20 critical rotating assets, run one technology well for six months, count the catches and the false alarms, then expand on evidence.

What does predictive maintenance pay?

The most credible public figures come from the U.S. Department of Energy's Federal Energy Management Program O&M guidance, maintained by Pacific Northwest National Laboratory:

Read the 8–12% honestly: it is the marginal gain over an already-functioning preventive program, and it assumes the program works, trained people, maintained baselines, alerts that become work orders. The larger 30–40% number belongs to plants climbing out of reactive maintenance, and most of that gain comes from planning discipline, not sensors.

Where should PdM fit in your strategy?

PdM is a rung on a ladder, not a destination for every asset. Non-critical equipment stays on time-based PM or deliberate run-to-failure. Critical assets with measurable degradation and expensive failures are the PdM candidates. The strategy-per-asset decision, and the honest cost of each rung, is the subject of our equipment reliability guide and the metrics that prove whether the investment worked (MTBF trending up, unplanned downtime trending down) live in your maintenance KPI set.

One last honest note: the hardest part of PdM in most plants is not analytics. It is that condition data, work orders, downtime logs, and parts inventory live in four disconnected systems, so nobody sees the whole picture. Connecting those sources into one layer, machines, software, and paperwork together, is the problem Harmony was built around; the CLS case study shows what unified plant data looks like in practice.