How to use the Weibull analysis calculator
- Enter life data, one row per unit: the time (hours, cycles, kilometres — any consistent unit) and a status, F for a failure or S for a suspension (a unit removed or still running without having failed). Paste from a spreadsheet, upload a .csv/.xlsx file or start from the template.
- Choose the method: median rank regression (the classic probability-plot method, recommended for small samples) or maximum likelihood (preferred with many suspensions; gives confidence bounds).
- Optionally enter a time t to get R(t), F(t) and the failure rate, and a custom B-life percentage.
- Read β, η, the probability plot, B-lives and curves. Print or save the report.
Analyse one failure mode at a time. Units that failed for a different reason are suspensions for the mode under study.
The Weibull distribution
The 2-parameter Weibull model gives the fraction failed by age t as F(t) = 1 − exp(−(t ⁄ η)β).
| Quantity | Formula |
|---|---|
| Reliability | R(t) = exp(−(t ⁄ η)β) |
| Failure (hazard) rate | h(t) = (β ⁄ η)(t ⁄ η)β−1 |
| Bx life (x% failed) | Bx = η·(−ln(1 − x ⁄ 100))1⁄β |
| Mean life | MTTF = η·Γ(1 + 1 ⁄ β) |
| Median life | B50 = η·(ln 2)1⁄β |
η, the characteristic life, is the age at which 63.2% of units have failed, whatever β is. The B10 life is the age by which 10% have failed (90% reliability) — the usual design-life figure for bearings and many mechanical components.
Interpreting the shape parameter β
| β | Failure rate | Typical meaning |
|---|---|---|
| β < 1 | Decreasing | Infant mortality: manufacturing and assembly defects, inadequate burn-in |
| β ≈ 1 | Constant | Random failures independent of age; the exponential model applies |
| 1 < β < 4 | Increasing | Early wear-out: fatigue, corrosion, erosion |
| β ≥ 4 | Rapidly increasing | Old-age wear-out concentrated in a narrow age band |
β drives maintenance policy: age-based replacement only pays when β is clearly above 1.
Median rank regression
Failures are sorted and each is given a plotting position. With suspensions, Johnson’s adjusted rank is used: adjusted rank = (reverse rank × previous adjusted rank + n + 1) ⁄ (reverse rank + 1). Benard’s approximation then gives the median rank F = (adjusted rank − 0.3) ⁄ (n + 0.4). Plotting y = ln(−ln(1 − F)) against x = ln t turns the Weibull CDF into a straight line with slope β. By default the tool regresses X on Y (time on rank), the convention for life data because the uncertainty is in the times; Y on X is available for comparison with spreadsheet solutions. R² measures how straight the plot is.
Maximum likelihood
MLE finds the β and η that maximise the likelihood of the observed failures and suspensions: failures contribute the density, suspensions the survival probability. The shape is the root of Σtᵢβ ln tᵢ ⁄ Σtᵢβ − 1 ⁄ β − mean(ln t of failures) = 0, and η = (Σtᵢβ ⁄ r)1⁄β, with sums over all units and r failures. Confidence bounds come from the Fisher information matrix (two-sided, at the level you choose) and are shown on the plot, for B-lives and for R(t). They are large-sample approximations and too narrow for a handful of failures. MLE also tends to overestimate β in small samples, which is why rank regression is commonly preferred below about 20 failures.
Worked example
Data: five units fail at 150, 290, 410, 560 and 800 hours (no suspensions).
Median ranks (Benard, n = 5): (i − 0.3) ⁄ 5.4 = 12.96%, 31.48%, 50.00%, 68.52%, 87.04%.
Transform: x = ln t = 5.011, 5.670, 6.016, 6.328, 6.685; y = ln(−ln(1 − F)) = −1.974, −0.973, −0.367, 0.145, 0.714.
Regression of x on y gives β ≈ 1.617 and η ≈ 515.6 hours with R² ≈ 0.9994.
B10 life = 515.6 × (−ln 0.9)1⁄1.617 ≈ 128 hours; MTTF = 515.6 × Γ(1.618) = 515.6 × 0.8957 ≈ 462 hours. Paste the five times into the tool to reproduce these values and see every intermediate number.
Exponential / MTBF mode
When the failure rate is constant (β = 1), only the total operating time T and the number of failures r matter: MTBF = T ⁄ r and R(t) = e−t⁄MTBF. Confidence bounds use the chi-square distribution: the lower bound is 2T ⁄ χ²1−α; 2r+2 for a time-terminated test (2r degrees of freedom if the test ended at a failure) and the two-sided upper bound is 2T ⁄ χ²α⁄2; 2r. With zero failures there is no point estimate, but the lower bound still exists — the basis of “demonstrated MTBF” tests. Quantiles are computed from the incomplete gamma function rather than a table.
Assumptions and limitations
- Two-parameter Weibull only; a failure-free period (3-parameter Weibull), mixed populations and interval-censored inspection data are not modelled.
- Suspensions are right-censored and non-informative (the reason for removal is unrelated to impending failure).
- Tied failure times each receive their own rank.
- Extrapolating far below the first failure (for example B0.1 from 10 failures) is highly uncertain.
If the plot is clearly curved, check other distributions — the normality test on ln t checks a lognormal model, and a histogram shows mixed populations. Reliability losses feed directly into availability in the OEE calculator. Methods follow Abernethy, The New Weibull Handbook, and Nelson, Applied Life Data Analysis.
Frequently asked questions
What do the Weibull shape and scale parameters mean?
The shape parameter beta describes how the failure rate changes with age: below 1 it falls (infant mortality), near 1 it is constant (random failures) and above 1 it rises (wear-out). The scale parameter eta, the characteristic life, is the age by which 63.2% of units have failed.
What is B10 life?
B10 life is the age by which 10% of a population is expected to have failed, so reliability at the B10 life is 90%. It is calculated as eta × (−ln 0.9)^(1/beta) and is widely used as the rating life of bearings and other mechanical components.
What is a suspension in Weibull analysis?
A suspension, or right-censored observation, is a unit that had not failed by the mode under study when it was removed from test or when the data were collected. Its running time is still valuable information and is included through adjusted ranks in rank regression or through the survival term in the likelihood.
Should I use median rank regression or maximum likelihood?
Median rank regression is the traditional probability-plot method and is generally preferred for small samples, roughly under 20 failures, because MLE tends to overestimate the shape parameter there. MLE is preferred for larger samples and heavily censored data, and it provides confidence bounds.
How is MTBF different from MTTF?
MTTF, mean time to failure, is the mean life of non-repairable items and equals eta × Γ(1 + 1/beta) for a Weibull distribution. MTBF, mean time between failures, applies to repairable systems with a constant failure rate and is the total operating time divided by the number of failures.
How do I calculate a confidence bound on MTBF?
For a constant failure rate, the lower confidence bound is 2T divided by the chi-square quantile at the chosen confidence with 2r + 2 degrees of freedom for a time-terminated test, or 2r degrees of freedom for a failure-terminated test, where T is the total operating time and r the number of failures.