What is Weibull analysis in maintenance reliability?
Weibull analysis models the statistical distribution of time or cycles to failure for a defined population and failure mode. In the two-parameter model, beta describes the shape of age-related hazard and eta is the characteristic life at which 63.2% of the modeled population has failed. Estimates should account for surviving or right-censored items, uncertainty, operating exposure and the suitability of the model before supporting maintenance policy changes.
Interpret Weibull parameters as properties of a fitted population, not asset identities.
The two-parameter Weibull reliability function is R(t) = exp[−(t/eta)^beta] for t greater than or equal to zero, beta greater than zero and eta greater than zero. Its hazard is h(t) = (beta/eta) × (t/eta)^(beta−1). Time t may be cycles or run hours where the units and population are consistent.
Beta below 1 describes decreasing modeled hazard with age; beta approximately 1 gives an approximately constant hazard; beta above 1 gives increasing hazard. Eta is the age where the modeled cumulative failure fraction is 63.2%, not the mean life, median life, or guaranteed replacement date.
| Parameter or result | Correct interpretation | Common misuse |
|---|---|---|
| Beta < 1 | Modeled hazard decreases with age | Automatically diagnosing installation errors without physical evidence |
| Beta = 1 | Constant modeled hazard; exponential special case | Declaring condition monitoring irrelevant |
| Beta > 1 | Hazard increases with age in the modeled population | Assuming preventive replacement must be economical or effective |
| Eta | 63.2th percentile of modeled failure times | Calling it mean time between failures or a mandated PM interval |
| R(t) | Modeled survival probability to specified age | Claiming it is the field-observed proportion without model checking |
Get complete age and censoring evidence before fitting a curve.
Define a consistent start-of-life or renewal event, mode-specific end-of-life definition, age in operating hours or cycles, and population inclusion rule. Keep running assets and items removed for other reasons as right-censored or otherwise appropriately treated observations rather than discarding them.
A motor still running after 2,000 hours contributes survival information even though it is not a failed item. An item removed for unrelated reasons may require censoring or competing-risk treatment depending on the question. Mixing installed ages, repair times, overhaul ages and system operating hours can invalidate the result.
- Asset serial number or equivalent observation ID and comparable duty.
- Failure mode, failure event and time or cycle of occurrence.
- For unfailed items, current exposure and censoring status.
- Replacement, overhaul and renewal definition consistent with the life model.
- Operating context, loading, maintenance policy and changes over the observation period.
- Uncertain or missing ages explicitly tagged, not invented.
Illustrative bearing population with beta 2 and eta 1,000 hours.
Assume, purely for demonstration, a validated two-parameter Weibull fit gave beta = 2 and eta = 1,000 operating hours for a comparable bearing failure mechanism. Reliability at 500 hours is exp[−(500/1000)^2] = exp(−0.25) ≈ 0.7788, or approximately 77.9% model survival to 500 hours. The cumulative modeled failure proportion by that age is approximately 22.1%.
This illustrates a survival calculation, not a fitted real-world bearing dataset. At 1,000 hours, model survival is exp(−1) ≈ 36.8%. That does not mean all bearings should be changed at 1,000 hours; optimization additionally requires replacement effectiveness, costs, consequence, uncertainty and conditional remaining life.
A population-based survival curve should not be presented as a precise remaining-life estimate for a particular bearing without asset-specific condition and duty evidence.
Estimate with censoring and test alternative interpretations.
Use validated reliability software for maximum-likelihood estimation with censoring when sample size and modeling assumptions support it. Report beta and eta with uncertainty bounds and fit diagnostics. The less data you have, the more seriously competing explanations and wide confidence intervals matter.
Separate distinct damage mechanisms before fitting. A mixture of random electrical failures, lubrication contamination, fatigue and installation errors may create a misleading aggregate hazard shape. Confirm whether field observations are truly individual first-failure or renewal lifetimes and whether changes in duty or redesign make the population nonstationary.
Repairable-system recurring-event behavior may require a recurrent-event, renewal or nonhomogeneous Poisson process treatment rather than naively fitting all times between repairs as independent first-life observations.
Increasing hazard is not the same as economically justified scheduled replacement.
Age-related hazard is one ingredient in a planned replacement policy. Also consider the fraction of failures the intervention would actually prevent, likelihood of infant mortality after replacement, predictive indicators, downtime consequence, replacement cost, operating access and mandatory engineering or safety requirements.
If the analysis shows no credible age relationship, do not manufacture one by changing the plot. If condition evidence allows reliable early intervention, RCM may still favor condition-based maintenance. The selected policy must be verified in the actual plant over an appropriate observation horizon.
What should Weibull analysis change?
It should make the uncertainty and nature of failure-age behavior explicit, challenge unsupported calendar replacement, and provide one defensible input to an RCM policy decision. It should not independently dictate a new replacement interval.
- Keep right-censored survivors and mode-specific exposure in the dataset.
- Distinguish increasing hazard from a cost-effective intervention.
- Verify model fit, uncertainty and mixed-population risks.
- Do not treat recurring repairable-system failures as independent first lifetimes without justification.
Technical references
- NIST/SEMATECH Engineering Statistics Handbook: Weibull — Two-parameter Weibull reliability and hazard definitions.
- NIST: Weibull MLE with Censored Data — Worked estimation example incorporating nonfailed life observations.
Use life data to challenge assumptions, not to decorate a PM interval.
Fit comparable, well-defined failure modes with survival information and integrate uncertainty into RCM decisions.
RCM learning path See facility-specific XRVE services