Preventive vs Predictive Maintenance Which Strategy Is Right for Your Factory

Preventive maintenance schedules work according to time, cycles or usage. Predictive maintenance uses measured condition and a model to estimate when intervention is needed. A factory should not choose one method for every machine. It should assign a strategy to each asset class according to criticality, failure behavior, detectability and economics.

The practical objective is not to maximize the amount of maintenance. It is to control risk at the lowest justified life-cycle cost. Some components need scheduled replacement. Some deserve continuous condition monitoring. Others can be allowed to fail because the consequence is small and recovery is quick.

What is the difference between preventive and predictive maintenance

Preventive maintenance triggers work at a planned interval. Predictive maintenance triggers a decision from evidence about current or future condition. Condition-based maintenance reacts to measured thresholds, while predictive methods estimate progression or remaining time.

Time-based preventive work is useful when deterioration relates reasonably to age or use, an inspection is required by law or the manufacturer defines a justified service interval. Examples can include lubrication, filter replacement, proof testing and inspections. The interval should be reviewed against failure history; repeating unnecessary work can introduce errors and consume production time.

Condition-based maintenance observes a variable such as vibration, temperature, oil condition, electrical signature or wear. Work begins when the condition crosses a defined limit or shows an abnormal pattern. It does not always predict a future failure date.

Predictive maintenance adds a forecast. The model may estimate failure probability, degradation state or remaining useful life. It can use physics, statistics, machine learning or a combination. The forecast must be accurate enough and early enough for the plant to act.

Run-to-failure is also a deliberate strategy when a component is inexpensive, non-critical, safe to replace after failure and held in stock. Treating every reactive event as a maintenance failure can lead to excessive prevention.

Strategy	Trigger	Best fit	Main risk
Run to failure	Functional failure	Low-consequence, replaceable and redundant items	Unrecognized dependency or safety impact
Preventive	Time, cycles or usage	Age-related wear, required inspections and stable tasks	Overmaintenance and introduced defects
Condition based	Measured condition or threshold	Detectable degradation with enough response time	Poor thresholds and nuisance alarms
Predictive	Forecasted condition or probability	High-consequence assets with useful data and actionable lead time	Model drift, false alerts and weak labels

Which maintenance strategy fits which assets

The correct strategy depends on consequence, failure mode, detectability, lead time, redundancy and the cost of monitoring. Asset price alone is not a sufficient measure of criticality.

Start with function. Record what the asset must do and what happens when it cannot. Consequences can include safety, environmental release, quality loss, lost constraint time, damage to other equipment, delayed orders and repair cost. A modest sensor on a bottleneck conveyor can be more critical than an expensive non-constraint machine.

Then examine failure behavior. Does wear increase with time or cycles? Is there a measurable degradation path? Does failure occur randomly with little warning? Can an inspection reveal the condition without stopping production? How much time exists between detection and functional failure?

Asset situation	Likely starting strategy	Reason	Evidence needed
Safety-critical protective device	Required proof test plus approved monitoring	Compliance and hidden failure risk	Legal, standard and engineering basis
Bottleneck rotating equipment with long repair lead time	Condition based or predictive plus planned tasks	High downtime consequence and measurable degradation	Failure history, vibration or other condition data
Low-cost redundant auxiliary fan	Run to failure or simple inspection	Low production consequence and quick replacement	Confirmed redundancy and stock
Wear component with stable cycle life	Usage-based preventive replacement	Failure relates to cycles	Replacement and failure distribution
Unique machine with rare unlabeled failures	Reliability improvement before AI	Data may be insufficient for prediction	Root-cause history and instrumentation plan
Large fleet of similar assets	Condition or predictive pilot	Repeated examples can support validation	Consistent tags, operating context and outcomes

The matrix should be reviewed by operations, maintenance, safety and finance. Operations understands production consequence. Maintenance understands failure and repair. Safety establishes non-negotiable controls. Finance helps value downtime and inventory. No single department holds the complete decision.

What do manufacturing studies say about maintenance performance

Research indicates that more advanced maintenance practices can be associated with better outcomes, but the evidence should not be turned into a universal promise. Sample, definitions and method matter.

A NIST survey received 85 responses and used 71 in analysis. Reported average maintenance shares were approximately 17.3 percent predictive, 31.8 percent preventive and 45.7 percent reactive. Plants with a high reactive-maintenance share were associated with much more downtime and more defects than less reactive groups. Depending on the grouping used, the more predictive group showed about 15 or 18.5 percent less downtime and substantially fewer defects.

These results are useful for direction, not a guaranteed saving. The sample was limited and the analysis does not prove that changing a maintenance label caused every difference. Better-managed plants may also have stronger data, training, planning and production systems. The full peer-reviewed maintenance study reports the method and limitations.

Older US Department of Energy guidance has often been quoted for figures such as 8 to 12 percent savings over preventive maintenance and 35 to 45 percent downtime reduction. Those values are historical estimates with methodological limitations. They should not appear as a current benchmark or a vendor guarantee.

The strongest conclusion is narrower: highly reactive operation carries risk, and companies should evaluate advanced maintenance from their own baseline. That is consistent with the NIST authors’ caution.

When does predictive maintenance make financial sense

Predictive maintenance makes financial sense when deterioration is detectable early enough, failure consequences are material and expected avoided loss exceeds the complete cost of data, models and response. The calculation must include false positives and missed failures.

Begin with the failure mode, not the sensor. Estimate annual event frequency and the consequence of each event. Consequence can include repair, secondary damage, lost contribution during downtime, expedited freight, scrap and restart loss. Use a range when history is sparse.

Then estimate detection performance within the required lead time. A model that identifies a problem five minutes before failure may be technically accurate but operationally useless when a replacement part takes two weeks. Measure precision, recall and lead time under real operating conditions.

Expected annual benefit can be expressed as failures correctly detected and acted upon multiplied by avoided consequence, minus false-alert cost and the cost of additional planned intervention. Subtract sensors, connectivity, software, model development, validation, internal labor, support and retraining.

Assume a critical asset has an expected failure loss of EUR 80,000 every two years, or EUR 40,000 per year before uncertainty. A condition program is expected to avoid half of that loss, but the plant can act on only 80 percent of useful warnings. Gross expected benefit is EUR 16,000 per year. If recurring monitoring and response cost EUR 12,000, the margin is narrow before implementation cost. The example shows why high failure cost alone does not guarantee an attractive project.

Use sensitivity. Test a lower event rate, worse detection, higher false-alert cost and delayed adoption. Predictive maintenance should not be approved with one precise forecast built from rare failures.

What data and capabilities are required

A predictive program requires reliable failure records, operating context, suitable condition signals, a validation method and a process for acting on alerts. More data does not repair ambiguous outcomes.

CMMS records should use consistent asset identifiers and failure codes. Work orders need symptom, failure mode, cause, action and result. Free text can add detail, but a model cannot learn reliably when the same failure appears under several names or completed work orders do not confirm what technicians found.

Condition data must include context. Vibration can change with speed, load, product and mounting. Temperature can change with ambient conditions. A model trained on one product may flag normal operation on another. Synchronize relevant operating state, maintenance actions and configuration changes.

Data ownership is ongoing. Someone must monitor missing signals, sensor replacement, model drift and changes to thresholds. Technicians need a way to confirm or reject alerts. Their feedback should update the record without being treated as infallible ground truth.

The IIoT connection should match the use case. InduVista’s industrial IoT implementation guide explains how to price the data path and security controls. A predictive pilot does not justify connecting every machine.

How should a predictive maintenance pilot be run

A predictive maintenance pilot should cover one critical asset class with a detectable failure mode, enough examples for validation and a clearly valued consequence. It should compare the proposed method with the current best alternative.

1. Select the failure mode and asset population.
2. Confirm that deterioration can be measured with actionable lead time.
3. Clean asset, work-order and operating-context records.
4. Establish the current event rate, downtime and maintenance cost.
5. Define precision, recall, lead-time and business acceptance criteria.
6. Run the new method in shadow mode before changing maintenance decisions.
7. Record true alerts, false alerts, missed events and technician findings.
8. Price the complete workflow, including inspections triggered by alerts.
9. Review safety and warranty requirements before changing scheduled work.
10. Scale only when technical and economic performance is repeatable.

A pilot may conclude that a simple threshold works as well as a predictive model. That is a useful result. It avoids paying for complexity that does not improve the decision.

How should a factory build a maintenance strategy portfolio

A maintenance portfolio should assign a documented strategy to asset classes, review actual performance and change the strategy when evidence changes. It should not become a one-time spreadsheet that is disconnected from work orders and production risk.

Start with the criticality and failure-mode analysis. Group similar assets only when function, duty, environment and consequence are comparable. For each significant failure mode, record the selected task, interval or condition signal, technical basis, owner and expected consequence. If no proactive task is technically effective, define run-to-failure controls, spares and recovery.

Connect the portfolio to planning. Preventive tasks should appear in the CMMS with the right materials, skills and instructions. Condition alerts need a response window, verification step and escalation. Predictive output should not create work automatically unless the risk and validation justify that control.

Review leading and lagging indicators. Leading measures include task compliance, overdue critical work, sensor coverage, alert response and data quality. Lagging measures include functional failures, downtime, repeat repairs, maintenance cost and quality loss. High preventive compliance with recurring failures can indicate that the tasks do not address the actual failure modes.

Use a controlled change process. When an interval is extended or a task is replaced by condition monitoring, record the engineering basis, safety or warranty constraints and the review date. Monitor the selected asset group before applying the change broadly.

At least annually, and after significant failures or process changes, review whether the asset remains in the right strategy class. New product loads, reduced redundancy, obsolete spares or a longer supplier lead time can change criticality even when the machine itself has not changed.

What do NIST researchers say about maintenance investment

Maintenance choices need a company-specific economic evaluation because asset condition, failure consequence and current practice differ.

“Individual businesses still need to evaluate investments in maintenance from their current circumstances and competitive strategies.”

Douglas S. Thomas and Brian A. Weiss wrote this in their NIST work on the economics of manufacturing machinery maintenance. The statement is especially important when broad industry percentages are used in sales material.

Frequently asked questions

Is condition-based maintenance the same as predictive maintenance

No. Condition-based maintenance can trigger work when a measured threshold is crossed. Predictive maintenance estimates future degradation, failure probability or remaining time.

Does predictive maintenance replace scheduled service

Not necessarily. Regulatory inspections, lubrication and age-related tasks may remain scheduled. Predictive methods should replace or change a task only when evidence and requirements support it.

Which machines should be monitored first

Start with assets that have high consequences, detectable degradation, actionable lead time and enough data to validate the method. Avoid selecting only by purchase price.

Is AI required for predictive maintenance

No. Thresholds, trend rules, physics and statistical models can be effective. AI is useful when it improves the decision and can be validated. See InduVista’s guide to AI in manufacturing.

Sources

1. Maintenance Costs and Advanced Maintenance Techniques in Manufacturing Machinery, NIST, 2021.
2. Peer-reviewed full text of the maintenance study, Journal of Research of NIST, 2021.
3. Economics of Manufacturing Machinery Maintenance, NIST AMS 100-34, 2018.
4. Operations and Maintenance Best Practices Guide, US Department of Energy, historical guidance.
5. ISO 17359:2018 Condition Monitoring and Diagnostics of Machines, ISO.
6. Challenges in Predictive Maintenance, CIRP Journal of Manufacturing Science and Technology, 2023.