故事 · "平均"两个字制造的最大误会
Origin · The Word "Mean" Built the Biggest Misconception
二战后军方与电子工业急需一个统一的"耐用度"指标,于是 MTBF 诞生:把一堆设备的总运行时间除以总故障次数,得到平均每隔多久坏一次。
它简洁、可加、好招标 —— 但也埋下了最深的误会。很多人把 "MTBF = 10000 小时" 直接读成 "能用一万小时不坏",这是错的。
在随机失效期(浴盆曲线平底段),失效是无记忆的:它服从指数分布,可靠度 R(t) = e−λt。
代入 t = MTBF = 1/λ,得 R = e−1 ≈ 37% —— 到平均寿命点时,已有六成多坏掉了。
MTBF 是失效率的倒数,不是"保用期";真正决定"用到某时刻还活着"的,是那条指数衰减曲线。
After World War II the military and the electronics industry needed a single "durability" number, and MTBF was born: total operating hours of a fleet divided by total failures — the mean time between two failures.
It was clean, additive, and easy to put on a procurement spec. It also planted the deepest misconception in reliability. Almost everyone reads "MTBF = 10,000 h" as "lasts ten thousand hours before breaking", and that is wrong.
On the flat bottom of the bathtub curve — the random-failure phase — failures are memoryless: lifetimes follow the exponential distribution, R(t) = e−λt.
Substitute t = MTBF = 1/λ and you always get R = e−1 ≈ 37% — by the average lifetime point, over 60% have failed.
MTBF is the reciprocal of a failure rate, not a warranty. What actually tells you whether a unit is still alive at a given instant is that exponential decay curve.
1 可靠度指数衰减:时间越久,活着的越少 Exponential Reliability Decay: Survivors Shrink with Time
R 0%2 浴盆曲线:MTBF 只在中间那段平底成立 The Bathtub Curve: MTBF Only Holds on the Flat Middle
失效率随寿命变化:早期(婴儿死亡)高且下降 → 随机失效期恒定平底(λ 不变,MTBF = 1/λ 在此成立) → 耗损期急剧上升。指数模型只适用于平底段。 Hazard varies across life: infant mortality starts high and falls → random-failure phase sits on a flat plateau (constant λ, where MTBF = 1/λ actually holds) → wear-out climbs steeply. The exponential model is only valid on the flat middle.
3 现实里的 MTBF MTBF in the Real World
硬盘 / 服务器:标称 MTBF 上百万小时,靠大批样本短期试验外推,绝非单台能跑那么久。
HDDs / servers: million-hour MTBF figures come from short tests on huge populations and extrapolation — no single drive runs that long.
MTBF vs MTTF:可修设备用 MTBF(故障间隔),不可修的一次性件用 MTTF(失效前时间)。
MTBF vs MTTF: repairable equipment uses MTBF (between failures); one-shot non-repairable parts use MTTF (time to failure).
任务可靠度:航空航天看 R(t) 而非 MTBF —— "执行 5 小时任务的成活率"才是真需求。
Mission reliability: aerospace specs the mission reliability R(t), not MTBF — "what's the survival probability over a 5-hour sortie?" is what actually matters.
恒定 λ 假设:MTBF=1/λ 只在随机失效期成立;进入耗损期,λ 飙升,再用就严重高估。
Constant-λ assumption: MTBF = 1/λ only holds while λ is constant. Once wear-out kicks in, λ climbs sharply and MTBF grossly overstates remaining life.
一句话In One Line
MTBF 的本质只是失效率 λ 的倒数 —— 单位时间坏得越慢,平均间隔越长。它是一个速率,不是寿命承诺。
真正回答"用到某时刻还活着没"的,是可靠度函数 R(t) = e−λt,它源于失效的无记忆性(指数分布)。
把 t = MTBF 代进去,永远得到 e−1 ≈ 37%:这正是"平均寿命点已有 63% 失效"的数学根源,与 MTBF 数值无关。
所以读 MTBF 时,记住三句话:它是速率不是寿命;到 MTBF 已折损六成;只在恒定 λ 的随机失效期才成立。
要谈"任务成活率",就直接看 R(t)。
MTBF is, at its core, just the inverse of the failure rate λ — slower failures per hour, longer mean interval. It is a rate, not a lifetime guarantee.
The honest answer to "is it still alive at time t?" is the reliability function R(t) = e−λt, which falls out of the memoryless property of exponential failures.
Plug t = MTBF and you always land on e−1 ≈ 37%. That is the mathematical root of "63% have already failed by the mean lifetime", and it does not depend on how big the MTBF number happens to be.
So when you read an MTBF, hold three lines in your head: it is a rate, not a lifetime; by t = MTBF you have already lost ~63%; it is only valid on the constant-λ random-failure phase.
If you actually care about mission survival, quote R(t), not MTBF.
常见误用Common Mistakes
把 MTBF 当保证寿命。到 MTBF 时 R ≈ 37%,已折损六成多;它是速率不是"用够多久才坏"。
Reading MTBF as a guaranteed lifetime. At t = MTBF, R ≈ 37% — over 60% have already failed. MTBF is a rate, not a "how long until it breaks" number.
耗损期还套 MTBF = 1/λ。公式只在 λ 恒定的随机失效期成立;老化阶段要用 Weibull 等模型。
Applying MTBF = 1/λ inside wear-out. The identity only holds while λ is constant. Once aging dominates, model the life with Weibull (or another wear-out distribution) instead.
把高 MTBF 等同高任务可靠度。任务可靠度看 R(t) = e−λt;同样 MTBF 下,任务越长成活率越低。
Equating big MTBF with high mission reliability. Mission survival is R(t) = e−λt. With the same MTBF, longer missions still drive R down — the rate doesn't care about how durable the headline sounds.