Why Weibull became the standard

Failure data is awkward: samples are small, some units haven't failed yet, and the failure behaviour changes shape from one mode to another. The Weibull distribution earns its place because one extra parameter lets the same curve describe infant mortality, random failure and wear-out – and because, plotted the right way, your data forms a straight line a plant engineer can read by eye.

Abernethy – a former Pratt & Whitney engineer who led the U.S. Air Force-sponsored work on the method – distilled four decades of practice into The New Weibull Handbook. It's not a statistics textbook; it's a field manual: which plotting positions to use, what to do with suspensions, how few failures you can get away with, when Weibull is the wrong tool. That practicality is why it's still cited, used and taught – in aerospace, mining, energy and manufacturing – nearly twenty years after the fifth edition.

Two parameters: β tells the story, η sets the clock

A two-parameter Weibull is fully described by the shape parameter β (beta) and the scale parameter η (eta, the characteristic life – the age by which 63.2% of units have failed). In plain terms: β tells you how the thing fails – dying young like a defective batch, at random like a puncture, or from old age like a tyre – and η tells you when, the age by which roughly two-thirds of them are gone.

FAILURE RATE → OPERATING TIME → Infant mortality β < 1 defects from manufacture, installation or maintenance Useful life β ≈ 1 random failures — external events, errors, overloads; age-independent Wear-out β > 1 fatigue, wear, corrosion — failure rate climbs with age the bathtub is the SUM of these three behaviours — an individual failure mode usually follows only one of them
Fig. 1 — The bathtub curve and where β lives, redrawn by Rob Reliability after Abernethy (2006). β < 1: failure rate decreasing (infant mortality). β = 1: constant (random; the exponential distribution). β > 1: increasing (wear-out). The β you fit identifies which regime your failure mode is in.

Abernethy's key teaching habit is to treat β as a diagnostic, not just a number. A fitted β around 0.6 isn't a statistic – it's an accusation: something in your supply chain, rebuild shop or commissioning process is shipping defects. A β of 3.4 on bearing failures with η far below catalogue life points at lubrication or alignment, not "old age". The parameter tells you where to investigate.

Reading a Weibull plot

On Weibull probability paper (log time vs. a double-log probability scale), Weibull data falls on a straight line. The slope of that line is β; the time where it crosses 63.2% is η. Don't let the scales intimidate you – the paper does the mathematics for you: if the points line up straight, Weibull fits, and you read both parameters directly off the chart.

99% 90% 63.2% 50% 30% 10% 5% 1% CUMULATIVE % FAILED (Weibull scale) 500 1,000 2,000 5,000 TIME TO FAILURE — hours (log scale) η ≈ 2,600 h (63.2% have failed by here) slope = β ≈ 2.1 steeper than 1 → wear-out behaviour
Fig. 2 — A schematic Weibull probability plot (illustrative data), drawn by Rob Reliability in the style of Abernethy (2006). Straight line = good Weibull fit. Slope = β. The fitted line crossing 63.2% gives η. A curved or kinked pattern is itself information – see the traps below.

Abernethy standardized the recipe most software now follows: rank the failures with median rank plotting positions (Benard's approximation), account for unfailed units as suspensions, prefer median rank regression for small samples and maximum likelihood for large ones, and always look at the plot before trusting the numbers. The straight line is not decoration – the eyeball test catches mixed failure modes and data problems that summary statistics hide.

From β to maintenance strategy

This is why a maintenance engineer should care about a statistics book: β tells you which maintenance policies can work at all.

Fitted βWhat it means physicallyWhat worksWhat doesn't
β < 1 Infant mortality – defects built in at manufacture, rebuild or installation. Find the defect source: precision installation, commissioning tests, burn-in, supplier quality. (See the I-P interval.) Scheduled replacement makes it worse – every intervention restarts the infant-mortality clock.
β ≈ 1 Random, age-independent failures – external events, process upsets, human error. Condition monitoring where a P-F window exists; root-cause work on the events; design out the stressor. Time-based replacement is pure waste – a new part is exactly as likely to fail tomorrow as the old one was.
1 < β < 4 Early wear-out – fatigue, corrosion, erosion, fouling. Condition-based maintenance first; age-based replacement only if CBM is impractical and the optimum interval beats run-to-failure on cost/risk. Ignoring age entirely; or replacing far too early "to be safe".
β > 4 Steep, predictable wear-out – the failure has a strong age signature (seals, liners, belts, some erosion modes). Scheduled restoration/replacement at an interval set from η, β and the cost ratio – this is where hard-time maintenance genuinely pays. Pure run-to-failure on critical assets – you're declining a cheap, reliable prediction.

Connect the dots

This table is the quantitative version of the six failure patterns in RCM II: patterns E and F are β ≤ 1, pattern B is high β. The aviation studies said most failure modes are not age-related – Weibull analysis is how you check what's true on your plant, per failure mode, instead of assuming.

B-lives: the number management understands

A B10 life is the age by which 10% of units are expected to have failed (B1 = 1%, B50 = median). Bearing catalogues have used B10 (L10) for a century; Abernethy generalized the habit to any component and any risk level.

B-lives turn a distribution into decisions:

  • Spare parts: if the B10 of a critical seal is 14 months and procurement lead time is 6, the min/max levels write themselves.
  • Replacement intervals: "replace at B5" is an explicit, defensible risk statement – unlike "every 2 years, because we always have".
  • Warranty & fleet decisions: comparing B10 before/after a design change, supplier switch or rebuild-shop change is the cleanest proof of improvement (Abernethy's "Weibull libraries" exist for exactly this).
  • Risk forecasts: with β, η and the fleet's age list you can forecast expected failures next quarter – the basis for shutdown scoping and budget defence.

Want to see it on your own data? Our Maintenance Strategy Optimization runs this end-to-end from CMMS exports.

The classic traps (Abernethy's own warnings)

Half the handbook is about not fooling yourself. The short version:

  1. Ignoring suspensions. Units still running are data. Fitting only the failures biases β and η badly – always include suspensions (censored units) in the ranking.
  2. Mixing failure modes. One Weibull per failure mode. A "pump failures" dataset mixing seal leaks and bearing seizures gives a meaningless β (often ≈ 1) and a curved plot. The cure is reading the work orders, not better math.
  3. Trusting tiny samples blindly. Weibull works impressively down to a handful of failures, but report confidence bounds – with 4 failures, β = 2.5 may not be distinguishable from β = 1.
  4. Wrong age basis. Calendar time vs. running hours vs. starts vs. tonnes – choose the clock the physics cares about. Erosion follows throughput, not the calendar.
  5. Forcing two parameters everywhere. A curved plot may signal a failure-free period: the three-parameter Weibull (t₀) handles it, but only with physical justification – not as a fit-improving cosmetic.
  6. The "garbage in" problem. CMMS dates are often work-order close dates, not failure dates. Clean the data before you fit it – the plot is only as honest as the history behind it.

References & further reading

This summary is original explanatory writing. All concepts belong to their authors – go to the sources.

  1. Abernethy, R.B. The New Weibull Handbook, 5th edition. Self-published, 2006. ISBN 978-0-9653062-3-2. Buy via Quanterion · Buy via Profit-Ability (Fulton)
  2. Weibull, W. "A Statistical Distribution Function of Wide Applicability." Journal of Applied Mechanics, 1951 – the original paper.
  3. ReliaSoft / weibull.com – free reference material on life data analysis. weibull.com
  4. Rob Reliability. Maintenance Strategy Optimization – how we run Weibull on client CMMS data, AI-assisted and engineer-verified.

Disclaimer. This page is an independent educational summary written entirely in Rob Reliability's own words. It is not affiliated with, sponsored by or endorsed by Dr. Robert B. Abernethy, his estate or his publisher. No text from The New Weibull Handbook is reproduced; the plots shown here are our own original illustrations built on illustrative data, and the underlying statistics (the Weibull distribution, the bathtub curve, B-lives) are part of the public mathematical and engineering literature. Book titles and trademarks remain the property of their respective owners and are used solely to identify the work being discussed. If you are a rights holder and have any concern about this page, contact us at hello@robreliability.com and we will address it promptly.

Done for you

Want this applied to your site?

We run Weibull on your CMMS exports at fleet scale: optimized PM intervals, over/under-maintenance flagged, savings quantified.

Maintenance Strategy Optimization
← All resources Run Weibull on your history