Weibull distribution: how to calculate reliability and MTBF

In one line

From a handful of real times between failures, the Weibull distribution calculates two parameters — Beta (shape) and Eta (scale) — that tell you whether equipment fails at random or is entering a wear-out phase, and from there you get the MTBF (mean time between failures).

What it is

The Weibull distribution describes the probability of failure of a piece of equipment as a function of time, with two parameters calculated from the real failure history:

• Beta (β), the shape parameter: defines how the failure rate changes over time.
• Eta (η), the scale parameter (also called Alpha): a characteristic life of the equipment, in the same units as the data (hours, cycles).

Beta's value is the one that matters most for maintenance, because it classifies which phase of the bathtub curve (or Davies curve) the equipment is in: when Beta is below 1, the phase is called infant mortality (decreasing failure rate); when Beta takes values close to 1, it is the useful life phase (constant, random failure rate); and when Beta is above 1, it is the aging or wear-out phase (increasing failure rate).

The Weibull fit also gives you the MTBF (mean time between failures), which is not a simple average of the observed times but is calculated with the Gamma function:

MTBF = η × Γ(1 + 1/β)

How Beta and Eta are calculated from real data

The classic method (Benard's median rank, with least-squares linear regression) needs no statistical software, just a calculator and these steps:

1. Sort the times between failures from smallest to largest and number them from 1 to n.
2. Calculate the median rank of each data point with Benard's approximation: RM = (i − 0.3) / (n + 0.4), where i is the data point's position in the sorted list.
3. Linearize the curve by calculating, for each data point: Y = Ln(Ln(1 / (1 − RM))) and X = Ln(the data point's time).
4. Run a linear regression of Y against X (least squares). That line's slope is directly Beta. The intercept (a) is used to get Eta: η = exp(−a / β).
5. Calculate the MTBF with the formula above, using the Gamma function evaluated at (1 + 1/β).

Real example

Mora documents the case of a machine with 6 recorded failures over a little more than a month and a half, with these times of operation without failures (TBF): 110, 330, 120, 220, 220, 220 hours.

Sorted from smallest to largest, with Benard's median rank calculated for each:

| Sorted data point (hours) | RM (Benard) | Y = Ln(Ln(1/(1−RM))) | X = Ln(hours) | |---|---|---|---| | 110 | 0.1094 | −2.1556 | 4.7005 | | 120 | 0.2656 | −1.1753 | 4.7875 | | 220 | 0.4219 | −0.6015 | 5.3936 | | 220 | 0.5781 | −0.1473 | 5.3936 | | 220 | 0.7344 | 0.2819 | 5.3936 | | 330 | 0.8906 | 0.7943 | 5.7991 |

The linear regression of those 6 points gives Beta = 2.3674, Eta = 234.18 hours and, with the formula above, MTBF ≈ 207.5 hours. That line's fit gives a correlation coefficient r ≈ 0.94.

But Mora himself points out a detail that changes the result: three of the six failures repeat the exact same time (220 hours). When there are repeated data points, the correct treatment is not to leave them as if they were three different points — they have to be collapsed into one (the most representative median rank, 0.7344 in this case) and the regression recalculated with the remaining 4 points instead of 6. With that correction, the result changes to Beta = 2.498, Eta = 220.60 hours and MTBF ≈ 195.7 hours, with a fit of r ≈ 0.96.

The difference is not cosmetic: the book itself sets an acceptable r² between 0.9025 and 1, and a valid r between 0.95 and 1. The naive calculation with all 6 points (r ≈ 0.94) falls just short of that threshold; the correction for repeated data (r ≈ 0.96) clears it. In both cases, Beta is clearly above 1: the machine in the example is in a wear-out phase, not failing at random.

What it is for

Knowing which phase of the bathtub curve a piece of equipment is in changes which maintenance strategy fits it, following the same logic laid out in RCM: with Beta below 1, it is worth investigating installation or start-up causes (infant mortality is not fixed with preventive maintenance); with Beta close to 1, failures are random and predictive maintenance has little to anticipate; with Beta above 1, as in the example, there is a real wear-out window where acting before the estimated MTBF prevents the failure.

The MTBF calculated this way is also useful for sizing critical spare parts and for objectively comparing the performance of similar pieces of equipment, instead of relying on perception or the last incident.

Benefits

• Turns a failure history into a concrete classification (infant mortality, useful life or wear-out) instead of a feeling that "this machine fails a lot"
• The MTBF it produces is statistically sounder than the simple average of the times between failures, because it accounts for the real shape of the distribution
• It can be calculated with a calculator and a table, with no need for specialized software

Limitations to keep in mind

• With few data points, the result is sensitive to details like the repeated values in the example — the same history gives Beta = 2.37 or Beta = 2.50 depending on how those ties are handled, and the MTBF difference is around 12 hours out of a total of about 200
• The fit is only trustworthy if the goodness-of-fit indices (r and r²) clear the minimum thresholds; a poor fit means Weibull is not the right distribution for that data, not that it should be forced anyway
• It assumes every failure in the history shares the same failure mode; mixing failure times from unrelated causes (say, an electrical failure and an unrelated mechanical one) into one dataset does not represent any real phenomenon

In summary

With nothing more than a piece of equipment's times between failures, Benard's median rank method lets you calculate Beta, Eta and the MTBF by hand. Beta is the key number: it tells you whether the equipment fails at random or is entering wear-out, and from there you know which maintenance strategy actually makes sense. Mora's example also shows that repeated data points are not a minor detail — handling them correctly or not is what decides whether the fit is statistically valid.