Weibull analysis: a complete, practical guide

Weibull analysis is the workhorse of reliability engineering: a method for taking a list of failure times and turning it into answers — how do these things fail, when will they fail, and what should we do about it? It has held that position for seventy years for one reason: the Weibull distribution is flexible enough to describe almost every failure behaviour that occurs in practice, using just two numbers.

This guide is the whole method, honestly told: what the parameters mean, how fitting actually works, the censoring issue that invalidates most quick analyses, a worked example with real numbers, and what to do with the results. Nothing here requires more than engineering-school statistics.

The distribution in one equation

The Weibull distribution says the probability that an item has failed by time t is:

F(t) = 1 − exp(−(t/η)^β)

Two parameters, each with a direct physical reading:

• β — the shape parameter. How things fail. It's the exponent on time, so it controls whether the failure rate falls, holds steady, or climbs as the item ages.
• η — the scale parameter, or characteristic life. When things fail: the time by which 63.2% of the population has failed, whatever β is. (Set t = η and F = 1 − 1/e ≈ 0.632 falls out of the equation.)

From F(t) come the other three functions you'll meet, all equivalent views of the same fit: the reliability (survival) function R(t) = 1 − F(t), the density f(t), and the hazard rate h(t) = f(t)/R(t) — the instantaneous failure rate of survivors, the curve maintenance decisions actually hinge on:

h(t) = (β/η) · (t/η)^(β−1)

β is the diagnosis

The reason engineers reach for Weibull rather than any other distribution is what β tells you before you've done anything else:

• β < 1 — infant mortality. The hazard falls with age: survivors are the good ones. Think manufacturing defects, bad installs, workmanship. Scheduled replacement is actively harmful here — you'd be swapping proven survivors for risky new units. Fix the process, screen with burn-in.
• β ≈ 1 — random failures. Constant hazard; age carries no information. This is the exponential distribution as a special case. Scheduled replacement buys nothing — run to failure, or monitor condition if the consequence is severe.
• β > 1 — wear-out. Hazard climbs with age: fatigue, wear, corrosion, erosion. Now age-based replacement can pay, and the steeper the β, the more predictable the failure and the stronger the case. β of 2–4 is typical mechanical wear-out; β above ~6 is so sharply timed you can almost schedule around it.

The three regimes are the three phases of the classic bathtub curve — but a single fitted β tells you which phase your failure mode actually lives in, which is worth more than the generic picture. One caution: a fleet mixing two failure modes (say, some infant mortality plus some wear-out) can fit to a meaningless in-between β. If the probability plot bends or the physics says two mechanisms, separate the modes and fit them individually.

How fitting works

Two families of methods, and knowing which to use matters more than most tutorials let on.

Median rank regression (MRR) is the classical, spreadsheet-friendly approach: sort the failures, assign each an estimated failure probability (Bernard's approximation), transform so a Weibull becomes a straight line, and fit the line — the slope is β. It's transparent, it pairs naturally with the probability plot, and you can do it in Excel in ten minutes.

Maximum likelihood estimation (MLE) asks instead: which (β, η) makes the observed data most probable? Each observation contributes what you actually know about it to a likelihood function, and an optimiser finds the parameters that maximise it. This is what modern software uses by default, for two hard-nosed reasons: it handles censored data correctly (next section — this is the decisive one), and its confidence bounds are well-founded rather than improvised.

On a clean, complete sample the two agree reasonably; on small or censored samples they can differ noticeably, and MLE is the defensible choice.

Censoring: the part that invalidates quick analyses

Everything above quietly assumed every unit ran until it failed. Real maintenance data is never like that — most of your fleet hasn't failed: units preventively replaced, still running, or removed for other reasons. These suspensions are right-censored observations, and they are not missing data. A bearing that ran 9,000 hours without failing is strong evidence about bearing life.

Leave suspensions out and your analysis studies only the weakest units in the population — the ones that failed — and generalises their frailty to the whole fleet. The bias is always pessimistic and routinely a factor of two on life estimates. The worked demonstration is here; the one-line summary is that MLE fixes it naturally (a failure at t contributes its density f(t); a suspension at t contributes its survival probability R(t)) and plotting methods don't.

The practical rule: if your data came from a CMMS, it has suspensions, and the analysis must include them.

A worked example

Eight seal failures, in operating hours:

410, 780, 1120, 1400, 1680, 1970, 2300, 2750

Fit a Weibull by MLE and you get β = 2.27, η = 1,753 hours. Reading it:

• β = 2.27 — wear-out. Seals degrade with age; age-based replacement is on the table. (The Excel walkthrough fits this same dataset by median rank regression and gets β ≈ 1.74 — a good illustration of how much method choice matters at n = 8.)
• Characteristic life 1,753 h: 63% of seals are gone by then.
• B10 life ≈ 650 h — failures start arriving in earnest around 650 hours. The median (B50) is ≈ 1,490 h and the mean ≈ 1,550 h — note how much later the "average" is than the number a planner actually needs.
• The hazard curve quantifies the wear-out: a survivor at 500 h is failing at ~0.26 per 1,000 h; at 2,000 h, ~1.5 per 1,000 h — six times the risk. That gradient is what justifies replacing old seals first.

Eight failures is a small sample, and the honest fit says so: confidence bounds on that B10 span roughly a factor of two. That width isn't a defect of the method — it's the true state of your knowledge, and it should shape how conservative the maintenance interval is until more data arrives.

Always look at the plot

A Weibull probability plot — the data on transformed axes where a true Weibull falls on a straight line — is the analysis's lie detector. Points hugging the line: the model fits. A curve or elbow: something's wrong, and the shape of wrong is diagnostic. A downward bow at early times often means a failure-free period (a three-parameter Weibull with a location offset, common in fatigue). A kink suggests two failure modes mixed together. Never report parameters from a plot you haven't looked at.

From parameters to decisions

The fit is the means, not the end:

• Choose the policy by β — run-to-failure for β ≈ 1, age-based replacement for wear-out, process fixes for infant mortality.
• Set the interval by economics, not eyeball: with the fitted model and the cost ratio of planned vs unplanned replacement, the cost-optimal preventive interval is a standard calculation (Reliafy runs it directly from any saved model).
• Set risk limits with B-lives — "interval ≤ B10" style targets for warranty, safety cases, and inspection planning.
• Feed system models: fitted component Weibulls are the inputs to reliability block diagrams and to MTBF/MTTF figures that actually mean something.

Frequently asked questions

What is Weibull analysis used for?

Fitting a life distribution to failure data in order to identify the failure pattern (infant mortality, random, or wear-out), estimate lifetimes with confidence bounds, and set maintenance, replacement, warranty, and spares policies. It's the standard first analysis in reliability engineering because the Weibull distribution can represent all three failure regimes with one shape parameter.

How many failures do I need for a Weibull analysis?

You can fit with as few as 2–3 failures plus suspensions, but expect very wide confidence bounds. Around 6–10 failures the picture firms up; beyond ~20 the parameters are usually stable. The right response to a small sample isn't to skip the analysis — it's to fit by MLE, report the bounds, and let their width temper the decision.

What does a Weibull shape parameter of exactly 1 mean?

β = 1 reduces the Weibull to the exponential distribution: constant hazard rate, failures arriving at random regardless of age. Preventive replacement has zero benefit — a new unit is exactly as likely to fail tomorrow as the old one it replaced.

What's the difference between the two-parameter and three-parameter Weibull?

The three-parameter form adds a location parameter γ: no failures can occur before time γ. It suits mechanisms with a genuine failure-free period, like fatigue crack initiation. Use it only when the physics supports it and the probability plot shows the characteristic early-time curvature — an unjustified γ is an easy way to flatter the fit.

Weibull analysis vs lognormal — which should I use?

Fit both and compare (probability plots plus a criterion like AIC). Weibull usually wins for wear-out and weakest-link mechanisms; lognormal often fits degradation processes that multiply, like crack growth. If they disagree materially in the region your decision lives in (say, the lower tail), that's a sign the data doesn't yet settle the question — widen the bounds you act on.