Weibull Analysis for Maintenance: Turning Failure Data Into Remaining-Useful-Life Predictions

Most plants collect years of failure data in their CMMS and never use it for anything beyond a mean-time-between-failures (MTBF) report. That's a waste. The same work-order history, analyzed with a Weibull distribution, tells you not just how often an asset fails on average, but when the next failure is likely — with confidence bounds you can plan around.

This is the statistical engine behind meaningful remaining-useful-life (RUL) prediction. No neural networks required.

What Weibull Analysis Actually Does

MTBF gives you a single number: the average time between failures. But failures don't arrive on a schedule — they arrive according to a pattern. Weibull analysis fits a distribution to your actual failure times and reveals that pattern.

The Weibull reliability function, in plain terms:

R(t) = exp( -(t / η)^β )

Two parameters do the work:

• η (eta) — the characteristic life. The age at which roughly 63.2% of the population has failed. A scale parameter.
• β (beta) — the shape parameter. This is the important one. It tells you what kind of failure mode you are dealing with.

The Shape Parameter Tells You the Story

β is where maintenance strategy lives:

• β < 1 — infant mortality / early failures. Failures cluster at the start of life. The asset is most likely to fail when it is new. Cause: manufacturing defects, bad install, wrong spec. Fix: better commissioning, burn-in testing, vendor QC — not more PM. Replacing these assets early makes things worse.
• β = 1 — random / exponential. Failures are time-independent. The asset is neither wearing in nor wearing out. Cause: external events — overloads, operator error, contamination, power spikes. Fix: condition monitoring and operator training. Time-based PM has no effect.
• β > 1 — wear-out. Failures accelerate with age. The older the asset, the higher the risk. Cause: fatigue, erosion, corrosion, bearing wear. This is the only regime where time-based preventive maintenance actually helps. Schedule replacement before the failure curve steepens.

A single number — β — tells you whether your PM schedule is helping, irrelevant, or actively harmful. Most plants never compute it.

Building a Weibull Analysis From Work-Order Data

You already have the inputs. You need:

1. Failure times. Time-between-failures for a given asset class (for example, "spindle bearings on CNC line 3"), measured in running hours or cycles — not calendar time if utilization varies.
2. Suspensions. Assets that have not failed yet (current runtime). These are censored data points and must be included — dropping them biases η upward and makes you overconfident.

With 5–10 failures plus suspensions, you can fit a credible 2-parameter Weibull. With 20+, you can attempt a 3-parameter fit (adds a failure-free interval γ) for assets with a clear break-in period.

The fitting steps

1. Rank failures (median rank: (i - 0.3) / (n + 0.4)).
2. Plot on Weibull probability paper, or fit via maximum likelihood — MLE is more robust with censored data.
3. Check the fit. If the points do not fall on a line, you likely have mixed failure modes — separate them by failure code and re-fit each. A Weibull plot that bends means you are modeling two different physical processes as one.

Reading the Output

Once fitted, the distribution gives you actionable numbers:

• B10 life — the age at which 10% of the population has failed. A common replacement target for critical assets.
• MTTF — mean time to failure: η · Γ(1 + 1/β).
• Failure rate at any age — the instantaneous risk. For β > 1, this rises steeply past η.
• RUL with confidence bounds — "given current age, the probability of surviving another X hours is Y%."

That last one is the payoff. Instead of "replace every 12 months," you get "this bearing is at the 70th percentile of its B10 life; failure probability in the next 30 days is 18% — schedule inspection now, replacement within 60 days."

Common Pitfalls

• Mixing failure modes. A bearing that fails from lubrication starvation (β ≈ 1) and one that fails from wear-out (β ≈ 2.5) cannot be modeled as one population. Segment by failure code first.
• Ignoring censored data. Suspensions carry real information. Excluding running assets inflates η.
• Calendar time vs. operating time. A standby pump run 200 hours a year is not the same age as one run 8,000 hours a year. Use operating hours.
• Too few data points. Below 5 failures, the fit is guesswork. Pool across identical assets or use a generic prior.
• Trusting the extrapolation. Weibull is reliable within the range of observed data. Extrapolating 5× beyond your last failure point is speculation.

How OpexMX Uses Weibull

OpexMX fits Weibull distributions to your work-order history automatically, per asset class, using maximum-likelihood estimation with censored-data handling. The output feeds each asset's health score and RUL estimate — so "remaining useful life" is not a marketing number, it is a probability with confidence bounds derived from your failure history.

See how OpexMX turns your CMMS data into failure predictions →

---

About OpexMX

OpexMX (Opex Maintenance eXecution System) is a cloud CMMS for manufacturing maintenance teams. OpexMX replaces WhatsApp-driven maintenance with structured work orders, workload balancing, asset history, and real-time dashboards. Built by Opex Consulting Group in Singapore, OpexMX is designed so technicians will actually use it. Learn more about OpexMX.