The Bathtub Curve and Weibull β: Reading Failure Behavior Over Life
The bathtub curve is the best-known picture in reliability engineering, and the Weibull shape parameter β is the number that tells you where on it a population sits. This guide covers both, along with the two most common ways β gets misread.
What reliability means
Reliability is the probability that a product performs its intended function, without failure, for a specified time under stated conditions. All four parts matter. A reliability number with no time attached, or with no operating conditions attached, is not a specification.
A plainer definition is also worth keeping in mind. Carl Carlson and Fred Schenkelberg, writing on the future of reliability engineering, point to the dictionary sense of the word: the quality of being able to be trusted to do what somebody needs. Customers don’t think in probabilities. They think in trust, and the probability is how engineers earn it.
The bathtub curve
A bathtub is a place to relax. Leave it to engineers to make a curve out of it. The bathtub curve plots the failure rate (hazard rate) of a population over its life. It is named for its shape: high at the start, low and flat through the middle, and rising again at the end.
- Infant mortality (decreasing rate). Early failures come from manufacturing and design defects that testing and quality control did not catch, such as weak solder joints, contamination or marginal parts. As the defective units fail and leave the population, the failure rate falls.
- Useful life (roughly constant rate). The product works as intended, and the failures that do occur are random, usually from external overstress events. The rate stays low and flat.
- Wear-out (increasing rate). Fatigue, wear, corrosion and other degradation mechanisms build up until failures rise steeply.
The bathtub is a composite. It is what you see when several failure modes act on one population, and real products often show only part of it. Much of today’s electronics shows little wear-out within its design life. Many mechanical parts never have a flat region, and start wearing out from the first hour.
What each region means for engineering decisions
| Region | What works | What does not |
|---|---|---|
| Infant mortality | Process control, design fixes for defect-prone features, and burn-in or HASS to screen escapes | Preventive replacement: a new part carries more early-life risk than the old one |
| Useful life | Design margin, derating, protection against overstress, and condition monitoring | Scheduled replacement: with a constant rate, a used unit is as good as a new one, so replacing it buys nothing |
| Wear-out | Preventive replacement before the rise, life extension through design, and planned end-of-life | Burn-in: it consumes life every unit needs |
Each region calls for a different response, so the first job with any field or test data is to find out which region the failures belong to. That is the job of the Weibull distribution.
Weibull β and η
The two-parameter Weibull distribution is the workhorse of life-data analysis because a single shape parameter lets it describe all three regions. Its reliability and hazard functions are:
- (beta) is the shape parameter. It sets how the failure rate changes with time.
- (eta) is the scale parameter, or characteristic life: the time by which 63.2% of the population has failed, whatever the value of β.
The exponent β − 1 in the hazard function is what links β to the bathtub:
- β < 1: the failure rate decreases with time, which is infant mortality.
- β = 1: the failure rate is constant. Weibull reduces to the exponential distribution and . This is the only case where a single FIT or MTBF number fully describes the population. See what 1 FIT really means.
- β > 1: the failure rate increases with time, which is wear-out. The larger β is, the more tightly failures cluster around η. Near β ≈ 3.5 the distribution looks roughly symmetric, like a normal distribution.
Two common misreadings of β
Misreading 1: “β measures how much the data varies.” β is a shape parameter. It is related to spread, since a higher β gives a narrower distribution relative to η, but its engineering meaning is the direction the failure rate moves over time: falling, flat or rising. Reading β as a scatter statistic throws away the most useful thing it says, which is which failure mechanism is at work.
Misreading 2: “Higher β means a more reliable product.” β alone cannot rank reliability. When wear-out happens is set by η. β only says how abruptly it arrives.
Take two populations with the same hours. Population A has and population B has .
| Time | R(t), A (β = 3) | R(t), B (β = 1.5) | More reliable |
|---|---|---|---|
| 1,000 h | 99.2% | 91.4% | A |
| 5,000 h (= η) | 36.8% | 36.8% | Equal |
| 8,000 h | 1.7% | 13.2% | B |
The steeper wear-out of population A protects it early in life and hurts it late. Which population is better depends entirely on the mission time. The reliability question is always “R at what time, under what conditions?”, the same as in the definition above.
One β per failure mode
A single two-parameter Weibull cannot draw the whole bathtub. Each region comes from different failure mechanisms, and each mechanism has its own β. Fitting one line through mixed failure modes produces a β that describes none of them. That usually shows up as a kink or S-shape on the probability plot. Split the data by failure mode, using failure analysis rather than eyeballing the plot, and fit each mode separately. Units that failed by another mode, or were removed from test unfailed, are treated as suspensions, not deleted.
From curve to decision
The bathtub curve is a mental model. β is the measurement behind it. Together they connect every stage of design for reliability: understanding the failure rate is what lets a team decide whether to fix the process, add margin, screen production or plan replacement, instead of applying all four everywhere.
Which region are your failures in? Enter failure and suspension times into the Weibull analysis calculator to get β and η with confidence bounds and a probability plot. A β whose confidence interval sits entirely below 1, entirely above it, or straddles it each points to a different action.
Related calculator
- Weibull Probability Plot
Fit 2-parameter Weibull to failure and suspension data; get beta, eta and B-lives.