Knowing a bearing is wearing out is useful, but knowing it has about sixty days left is far more valuable. Prognostic models turn condition data into a time estimate that planners can put straight into the maintenance schedule.
Predictive maintenance moves from spotting faults to forecasting when they will stop a machine. This guide explains the three main estimation approaches and shows a simple linear degradation calculation.
What Is Remaining Useful Life?
Remaining useful life, or RUL, is the estimated time or number of cycles a component can keep operating before it reaches a failure threshold or needs replacement. It is the forecasting step of predictive maintenance, going beyond simply detecting that a fault exists.
An RUL estimate is never a single exact number. Good models give a most likely value plus a confidence range that narrows as more data arrives.
MathWorks explains that RUL estimates help schedule maintenance, order spare parts and avoid unplanned downtime. The model choice depends on what data you actually have.
3 Proven Estimation Methods
Use lifetime data from many similar units.
Match a unit against run to failure histories.
Extrapolate a condition indicator to a known threshold.
MathWorks notes that degradation models predict when the indicator will cross the threshold, with confidence bounds on that time. Survival models suit cases where only failure times are recorded.
Similarity models need many complete run to failure records, which are rare in plants. They work well for large fleets such as engines or pumps.
Choosing by Available Data
| Data You Have | Suitable Model | Example |
|---|---|---|
| Only failure times | Survival model | Pump seal lives from work orders |
| Full run to failure histories | Similarity model | Aircraft engine fleet data |
| Trend plus known limit | Degradation model | Vibration against ISO limit |
| Live sensor streams | Machine learning on features | Motor current and temperature |
Condition indicators often come from vibration analysis with AI. Long histories are stored in a process historian, which is the natural data source.
From Sensor Data to a Forecast
Updating the model with each new reading narrows the bounds. A digital twin can add physics knowledge when data is thin.
Where the model runs affects response time and cost, see edge AI vs cloud AI. Simple trend models can run on an edge gateway.
The NASA CMAPSS Benchmark
A benchmark matters because prognostic methods are hard to compare on private plant data. Shared data lets researchers test ideas on the same engines, sensors and failure modes.
NASA released the CMAPSS turbofan engine simulation data, which became the standard benchmark for prognostics research. Ramasso and Saxena reviewed the many algorithmic approaches developed on it.
Their review is useful for students because it compares methods on the same data. It also shows how results depend on data preparation and scoring choices.
Linear Degradation Formula
Example, pump bearing vibration:
Alarm threshold 7.1 mm/s
Current value 4.1 mm/s
Trend rate 0.05 mm/s per day
RUL = 3.0 ÷ 0.05 = 60 days
About 8.6 weeks from today
Real degradation often accelerates near the end, so a linear fit can be optimistic. Recalculate the rate weekly and watch for curvature.
RUL Calculator
Add a safety margin to the result when parts have long lead times. Use it to book the repair during a planned stop.
- Plans maintenance before failure.
- Orders spares at the right time.
- Avoids unplanned downtime.
- Extends use of healthy parts.
- Needs good quality history data.
- Uncertainty can be wide early on.
- Linear trends miss sudden faults.
- Thresholds must be well defined.
Sudden failures such as those in battery failures or servo motor failures may give little warning. Watch for early signs listed in temperature sensor warning signs.
The broader methods behind these models are covered in AI and machine learning and AI in PLC, SCADA and DCS.
CMAPSS Prognostics Review PDF
Prognostics Video
Remaining Useful Life FAQ
Related Articles
- Predictive vs Preventive Maintenance
- Vibration Analysis With AI
- What Is a Process Historian
- Digital Twin in Industrial Automation
- Edge AI vs Cloud AI
External References
- Prognostics on CMAPSS Dataset, NASA
- Three Ways to Estimate Remaining Useful Life, MathWorks
- Prognostics, Wikipedia
What We Learn Today
- Remaining useful life forecasts time left before a failure threshold.
- Survival, similarity and degradation models suit different data.
- Linear trends give quick estimates but need frequent updates.