Weibull analysis is a method that fits a statistical distribution to a set of failure times and returns two numbers that describe how a part fails: beta (β), the shape parameter which tells you whether failures are early, random, or wear-out; and eta (η), the characteristic life the age by which 63.2% of the population has failed. Beta answers what kind of failure; eta answers when.
It is the most useful tool in a reliability engineer's kit because it turns a scatter of failure dates into a diagnosis and a maintenance decision. Named for Waloddi Weibull, who published the distribution in 1951, it underpins the bathtub curve and gives a rigorous alternative to the averages behind MTBF. This guide covers what beta and eta mean, how to plot real failure data, and how to act on the result.
What does the beta shape parameter tell you?
Beta tells you which region of the bathtub curve a part lives in, which is the single most important thing to know before you change a maintenance strategy. The failure rate over age moves with beta:
- β < 1, infant mortality. The failure rate falls as parts age. Survivors get more reliable, not less. This is the signature of manufacturing defects, bad installation, contamination at start-up, or maintenance-induced problems. Time-based replacement makes this worse you keep swapping good parts for fresh parts that then fail young.
- β ≈ 1, random failure. The failure rate is roughly constant; a part is no more likely to fail tomorrow than it was last year. This is the exponential distribution, and it describes failures with no age memory, random overloads, foreign objects, stray voltage. Scheduled overhaul does nothing here; you need condition monitoring or design changes.
- β > 1, wear-out. The failure rate climbs with age. Fatigue, erosion, corrosion, and abrasion all live here. This is the one region where planned, time-based replacement pays off, because there is a knee in the life where risk starts rising steeply. A β around 3 to 4 produces a nearly bell-shaped life distribution.
What is eta, the characteristic life?
Eta is the age at which 63.2% of the population will have failed, regardless of beta. It is the natural scale of the distribution, set the age equal to eta and the math always returns the same 63.2% failure fraction, which makes it a clean, shape-independent way to say how long the part lasts.
Two practical points. Eta is not the average life and it is not the median; it sits at that fixed 63.2% point because that is where the distribution's exponent equals one. And engineers rarely design to eta, they design to a B-life the age by which a chosen small fraction has failed. B10 life, for example, is the age by which 10% have failed, the standard bearing-rating convention. Report eta and beta together: eta without beta hides whether that life is a sharp cliff or a long gentle tail.
How do you plot failure data?
You rank the failure times, assign each a plotting position, and put them on Weibull probability paper. If the points fall on a straight line, the data is Weibull, the slope of the line is beta, and the age at the 63.2% gridline is eta. The steps are mechanical:
- List every failure time and every survivor. Record the age at failure for failed units and the current age for units still running. Those survivors are censored data, they have not failed yet, and throwing them out biases the result pessimistic.
- Rank the failures and assign median ranks. Sort failure times smallest to largest and give each an estimated cumulative failure percentage using median-rank tables or Bernard's approximation, adjusting the ranks to account for the censored survivors.
- Plot each failure point. Age on the horizontal log axis, cumulative percent failed on the vertical Weibull axis. Modern reliability software does this for you, but the plot is the point, it shows the fit, not just the numbers.
- Fit the line and read beta and eta. The best-fit slope is beta; drop a line from the 63.2% gridline to the age axis to read eta. Watch the fit itself: a clean line means one failure mode, a dogleg usually means two modes mixed together.
- Sanity-check the fit before you trust it. A curved or kinked plot means the data is not a single Weibull, often a sign that two failure modes are tangled and need to be separated and fitted on their own.
What does a worked example look like?
Take a mechanical seal on a process pump. Over two years, six seals failed at these operating ages: 900, 1,500, 2,100, 2,600, 3,200, and 3,900 hours. Four more seals of the same type are still running, censored survivors that entered service later. Fit the ten data points and the plot returns a slope of about β ≈ 2.3 and a characteristic life of roughly η ≈ 3,400 hours.
Read it: β above 2 is a clear wear-out pattern, so this seal ages predictably and a time-based interval is defensible. But η is not the interval, 63.2% of seals fail by 3,400 hours, which is far too many. The useful number is the B10 life, the age by which only 10% have failed. For this fit that lands near 1,300 hours. Replace on a roughly 1,200-hour interval and you head off most failures while still using the bulk of each seal's life. Ignore the fit and run to failure, and you are gambling with a part whose risk climbs every shift.
How do you act on a Weibull result?
The result changes the maintenance strategy, and this is where Weibull earns its keep over a blunt average. Map beta to an action:
| Beta | Failure pattern | What causes it | Right response |
|---|---|---|---|
| β < 1 | Infant mortality | Defects, bad install, contamination, poor rebuilds | Fix quality and installation; burn-in; stop time-based swaps |
| β ≈ 1 | Random | Overloads, foreign objects, stray events | Condition monitoring; design out the cause; no scheduled overhaul |
| β 1.5–3 | Early wear-out | Erosion, fatigue starting | Condition-based triggers; watch the trend closely |
| β > 3 | Rapid wear-out | Predictable end-of-life aging | Time-based replacement just before the knee in the curve |
The most common expensive mistake in maintenance is applying a fixed replacement interval to a part with β below 1, the classic finding that many overhauls introduce more failures than they prevent. Weibull is how you catch that before it costs you. Once you know a part is a genuine wear-out item, you can set the interval on your preventive maintenance schedule at a defensible age rather than a guess, and route the random and infant-mortality items to predictive maintenance or a root-cause fix instead. It all depends on failure records clean enough to fit, the same clean history that powers real equipment reliability work.
What the numbers say
- The 63.2% characteristic-life point and the beta interpretation are not conventions you can bend, they fall out of the distribution's math. The U.S. National Institute of Standards and Technology documents both in its reliability handbook: the shape parameter governs whether the failure rate decreases, stays constant, or increases, and the scale parameter marks the characteristic life (NIST/SEMATECH e-Handbook of Statistical Methods, §8.1.6.2 Weibull).
- Getting the strategy right pays. The U.S. Department of Energy's FEMP O&M guidance, maintained by PNNL, reports condition-based programs save 8–12% over preventive-only maintenance and 30–40% versus reactive operation (PNNL, O&M Best Practices: Maintenance Approaches). Weibull is how you decide which parts belong in which program instead of over-maintaining everything.
Weibull turns a pile of failure dates into two numbers and a decision. Beta names the failure pattern; eta scales the life; the plot shows whether you have one failure mode or two. The hard part is never the math, every reliability package and even a spreadsheet will fit the curve for you. It is having failure records clean and complete enough to fit, with real ages and honest censoring, which is exactly the floor-data problem Harmony was built to solve. See how one plant got there in the CLS case study.