Predictive vs. Preventive Maintenance: What Actually Matters in Practice
By Sync Motion — industrial software and OT/IT integration from Austria. We build real-time monitoring into existing industrial environments — from sensor layer to operational dashboards — without replacing what works.
Every plant has a maintenance strategy. In most cases, it's a mix: calendar-based intervals for what the manufacturer specifies, reactive repair for what breaks anyway, and a growing number of vendors promising that "AI-powered predictive maintenance" will solve the rest.
This article covers what predictive and preventive maintenance actually mean in technical terms, where each approach makes sense, and what it takes to move from calendar-based to condition-based maintenance in a real plant environment.
What the terms actually mean
They're used often, rarely defined cleanly.
Preventive Maintenance is time-based or cycle-based. Every 5,000 operating hours, once per quarter, per manufacturer specification. Components are replaced, lubricated, inspected — regardless of their actual condition. The basis is MTBF data, statistical averages, and operating experience. No sensors required, no real-time signals. The logic: if a bearing fails on average after 8,000 hours, replace it at 6,000. This is sound engineering, it's worked for decades, and it's the default in most plants worldwide.
Predictive Maintenance is condition-based. Sensors continuously or periodically measure the actual state of a component — vibration, temperature, motor current, oil particles, ultrasonics. When the data shows degradation, you intervene. Not before, not after. The goal is to time the intervention as closely as possible to the actual need, ideally within the P-F interval, the window between the first detectable fault signal and functional failure.
This can be simple: vibration exceeds threshold for a defined duration, alert goes out. Or complex: a model calculates remaining useful life based on spectral analysis, temperature trends, and operating profile. ISO 13374 describes the framework — from data acquisition through state detection to decision support.
The difference isn't academic. It determines whether you replace a bearing by calendar or by actual wear. Both are maintenance. The decision basis is different.
Where each approach makes sense
The question isn't "predictive or preventive." It's: which assets justify which approach?
Preventive maintenance works well for components with predictable, statistically narrow wear patterns. Filters, belts, seals — parts where the manufacturer's interval is reliable enough. Also for assets with low downtime risk, components without measurable early warning signals, and anywhere sensor retrofitting is disproportionately expensive or impractical. Regulatory-mandated inspection intervals — pressure vessels, safety-critical systems — stay calendar-based regardless of condition.
Predictive maintenance works well for assets where failure is expensive and degradation produces measurable early signals. Rotating equipment — motors, gearboxes, compressors, turbines — are the classic candidates. Bearing wear shows up in vibration data weeks before functional failure. Motor current analysis detects developing electrical faults. Pumps show cavitation through a specific frequency profile. When downtime costs $10,000 to $100,000+ per hour, a single avoided unplanned outage pays for the instrumentation across the entire plant.
In practice, the answer is almost always hybrid. Preventive as the baseline for 80% of assets. Predictive on the critical 20% that drive the majority of downtime costs. Starting with "everything predictive" fails on complexity and cost. Starting with the five most expensive failure sources delivers measurable results within a year.
What the numbers say
The data on predictive maintenance is solid at this point. The most widely cited results come from McKinsey, Deloitte, and PwC — and they're consistent across sources.
Maintenance costs: 18 to 25% reduction compared to calendar-based or reactive maintenance (McKinsey, "Prediction at Scale"). Unplanned downtime: 30 to 50% reduction. Asset life: 20 to 40% extension. Asset availability: 5 to 15% increase.
PwC, in a European study of 268 companies across Germany, the Netherlands, and Belgium, reports more conservative but realistic figures: 12% cost reduction, 9% availability increase. These are results from running operations, not laboratory settings.
Deloitte puts the total cost of unplanned downtime in US industry at roughly $50 billion annually. At the plant level, outage costs range from $50,000 to $260,000 per hour depending on the operation (ABB/Senseye). Payback periods for predictive maintenance implementations are typically under two years, with 95% of adopters reporting positive ROI and over a quarter achieving full payback in year one.
Important context: these numbers come from operations that approach this systematically. The reality is that 60 to 70% of PdM initiatives miss their ROI targets in the first 18 months, not because of the technology, but because of data quality, missing integration, and insufficient change management. The technology is not the problem. The rollout is.
What predictive maintenance actually requires
Here's where it gets concrete. Requirements differ significantly depending on whether you're working with simple threshold monitoring or ML-based prognostics.
Tier 1: Condition monitoring with thresholds — the entry point that delivers 60 to 70% of the value at a fraction of the cost. What you need: vibration sensors (accelerometers), temperature probes, motor current analysis, optionally ultrasonics and oil particle counters. IoT gateways or edge devices transporting data via OPC UA or MQTT to a historian or cloud system. A simple rule engine — vibration above limit for defined duration triggers alert — and integration with the CMMS for automatic work order generation. Investment: $5,000 to $50,000 per critical asset depending on sensor requirements and connectivity.
Tier 2: ML-based prognostics (RUL, anomaly detection) — for complex assets where simple thresholds aren't enough. This requires high-frequency data (kHz range for vibration analysis), edge preprocessing, an ML platform (Azure IoT, AWS, MindSphere, or on-premise), and models trained on the specific failure modes of the equipment. That requires either in-house data scientists or a managed service. Typical platform costs: $50,000 to $500,000 initial plus ongoing subscription.
What vendors don't emphasize: usable models need 6 to 24 months of operating data. A model trained on one asset doesn't automatically transfer to another. And when processes change, models drift — without ongoing maintenance, accuracy degrades. Starting with a pilot on 5 to 10 critical assets is realistic. Starting with an enterprise platform for the entire plant builds infrastructure before solving a problem.
A practical benchmark: simple threshold logic delivers 70 to 80% of the value at 20% of the cost. Full ML prognostics only pay off above a certain level of asset complexity and downtime cost. The right question isn't "AI or not" — it's "Where does an unplanned outage cost me enough to justify the investment?"
Common mistakes in implementation
A few points that come up repeatedly in predictive maintenance rollouts:
Starting with the platform instead of the problem. Which asset fails, what does it cost, and where is the signal measurable? That's the right sequence.
Scoping the pilot too large. 5 to 10 critical assets are enough to start. When the first alert prevents an unplanned outage, the business case for the rest of the plant is made.
Skipping condition monitoring. Many plants don't have continuous condition data yet and jump straight to "AI-powered prognostics." Without a clean data foundation, any model is guesswork.
Underestimating integration with existing systems. A vibration alert that doesn't land in a CMMS and doesn't generate a work order is a notification, not a maintenance process.
Not involving the maintenance team. Predictive maintenance changes workflows. If the technician doesn't understand what the system reports and why, they'll ignore it, rightly so.
The security angle
For plants in the EU: the NIS2 directive (EU 2022/2555) isn't a maintenance standard. It has no direct connection to maintenance strategy. But it has an indirect one that matters in practice.
Every predictive maintenance implementation expands the OT/IT attack surface. Sensors, gateways, cloud analytics, CMMS integration — these are new connection points between the plant network and IT. For companies in the 18 sectors covered by NIS2 — energy, certain manufacturing categories, transport — risk analysis, network segmentation, encryption, and incident reporting obligations apply. Including for new digital maintenance systems. Non-compliance penalties reach up to 2% of global annual turnover.
In practice, this means: anyone designing a PdM architecture today should build NIS2 compliance in from the start. Edge processing over unsecured cloud connectivity. Zero-trust architectures. OT network segmentation. This isn't overregulation — it's the reality of digitizing production data.
What stays
Preventive maintenance isn't going anywhere. For the majority of plant assets, calendar-based maintenance remains the right strategy. What changes is the addition: the critical assets that drive the bulk of downtime costs deserve condition-based monitoring. Not because it's modern, but because a single avoided unplanned outage pays for itself.
The path there isn't a technology purchase. It starts with a clear assessment: What fails? What does it cost? Where is the signal? And only then: what sensors, what platform, what integration.
Next steps
Sync Motion builds condition-based maintenance into existing industrial environments — from sensor layer to operational dashboards. PlantWatch, our monitoring platform, is designed for brownfield integration: connecting to what's already there, adding the visibility layer that calendar-based maintenance alone can't provide.
A technical conversation about your plant is always a good starting point.
office@sync-motion.com · sync-motion.com