故事 · 被嫌"太灵活"的天才分布
Origin Story · The Distribution Rejected for Being "Too Flexible"
1951 年,瑞典工程师 Waloddi Weibull 发表《一个广泛适用的统计分布函数》。审稿人嫌它"太灵活,什么数据都能拟合",一度不被看好。
可正是这份灵活,让它能用一个形状参数 k 同时描述材料疲劳、轴承寿命、风速、人类寿命。它的密度是
f(t) = (k/λ)(t/λ)k−1 e−(t/λ)k,失效率是 h(t) = (k/λ)(t/λ)k−1 ——
k 小于 1 失效率递减、等于 1 恒定(就是指数)、大于 1 递增。三段拼起来,就是可靠性工程里那条著名的浴盆曲线。
In 1951, Swedish engineer Waloddi Weibull published "A Statistical Distribution Function of Wide Applicability." Reviewers grumbled that it was "too flexible — it can fit anything", and the paper was initially dismissed.
That very flexibility is what lets a single shape parameter k describe material fatigue, bearing life, wind speed, and human lifespan, all from one family. Its density is
f(t) = (k/λ)(t/λ)k−1 e−(t/λ)k, and its hazard rate is h(t) = (k/λ)(t/λ)k−1 —
decreasing when k < 1, constant when k = 1 (the exponential), increasing when k > 1. Stack the three regimes and you get reliability engineering's iconic bathtub curve.
1 失效率 h(t):你现在站在浴盆的哪一段? Hazard Rate h(t): Which Stretch of the Bathtub Are You On?
随机期Random2 寿命分布 f(t):产品大多在什么时候坏 Lifetime Distribution f(t): When Most Units Actually Fail
k<1 时大量产品很早就坏(右偏严重);k 越大,失效越集中在特征寿命 λ 附近,曲线越像正态钟形。 When k < 1, many units die very early (heavy right skew). As k grows, failures bunch around the characteristic life λ and the curve approaches a normal bell.
3 现实里的威布尔 Weibull in the Real World
浴盆曲线:早夭、随机、耗损三段拼成一条 h(t),是可靠性工程描述全生命周期的通用语言。
Bathtub curve: infant mortality + random + wear-out, stitched into one h(t) — the universal language reliability engineers use for full life cycle behavior.
维修策略:k>1 进入耗损期 → 在失效率陡升前定期更换(轴承、刀具、轮胎)。
Maintenance strategy: once k > 1 (wear-out), schedule preventive replacement before the hazard rate spikes — bearings, cutting tools, tires.
寿命试验:用威布尔拟合寿命数据估 B10 寿命(10% 失效点),给质保与保修期定指标。
Life testing: fit Weibull to lifetime data and read off the B10 life (10% failure point) — the standard input for setting warranty terms and reliability targets.
早夭筛选:k<1 的婴儿死亡 → 出厂前老化烧机(burn-in),把缺陷品挡在保修期外。
Infant-mortality screening: when k < 1, weed out latent defects with factory burn-in so the bad units fail before they ship — not in customers' hands.
一句话In One Line
指数分布只能描述"不会老化"的随机失效(k=1 这一种)。威布尔把它推广:一个 k 就在早夭、随机、耗损之间自由切换 ——
这就是它成为可靠性工程地基的原因。看懂浴盆曲线,你就能回答最实际的问题:这东西该什么时候换?保修期定多久?要不要做老化筛选?
The exponential only describes the no-aging random-failure case (k = 1, one slice of the bigger picture). Weibull generalises it: a single k slides freely between infant mortality, random failure, and wear-out —
which is exactly why it became the foundation of reliability engineering. Read the bathtub curve and you can answer the questions that actually matter on the floor: when should we replace this part? How long should the warranty be? Do we need burn-in?
常见误用Common Mistakes
不看 k 直接套指数。先拟合 k:磨损件 k>1,绝不能当指数。
Defaulting to the exponential without checking k. Estimate k first — wear-out parts have k > 1 and the exponential is simply wrong for them.
样本太少硬估 k。小样本 k 的置信区间很宽,需足够失效数据。
Estimating k from a handful of failures. Small samples leave a wide confidence interval on k — wait until you have enough observed failures, or use a Bayesian prior.
把 λ 当成平均寿命。λ 是 63.2% 失效点(特征寿命),平均寿命另算。
Reading λ as the mean lifetime. λ is the 63.2% failure point (the characteristic life). The mean is λ · Γ(1 + 1/k) — a different number.