Visual walkthrough — Reliability — MTTF, MTBF, exponential failure model

3.6.31 · D2 · Physics › Spacecraft Structures & Systems Engineering › Reliability — MTTF, MTBF, exponential failure model

Yeh puri kahani hai exponential failure model ki, pictures mein batai gayi. Hum bilkul scratch se shuru karte hain — koi formulas assume nahi — aur ek-ek picture ke saath build karte hain, jab tak famous curve $R(t)=e^{-\lambda t}$ khud nahi nikal aata. Phir hum us curve se hi average lifetime (MTTF) bhi nikal lete hain.

Agar tumne pehle kabhi integral, derivative, ya letter $e$ nahi dekha — achha hai. Hum yahan har cheez ko kama ke laate hain.

---

Step 1 — Hum actually kya measure kar rahe hain: survival fraction

KYA. Socho hum 1000 identical components banate hain aur sab ko ek saath on karte hain. Jaise-jaise time guzarta hai, woh ek-ek karke mar jaate hain. Kisi bhi moment hum count karte hain kitne abhi bhi zinda hain aur 1000 se divide karte hain jisse shuru kiya tha. Woh fraction hi hamaari reliability $R(t)$ hai.

KYO. Ek single part ke liye probability slippery hoti hai — ya toh kaam karta hai ya nahi. Lekin ek badi bheed ka fraction ek real, drawable, measurable number hota hai. "Survival ki probability" word ke peeche yahi honest picture hai.

PICTURE. Figure mein, blue dots zinda hain, red dots dead hain. Shuru mein almost sab blue hain; baad mein, zyada red ho jaate hain. Time $t$ par blue curve ki height hi $R(t)$ hai.

---

Step 2 — Ek assumption: ek constant death rate $\lambda$

KYA. Hum ab ek modelling choice karte hain. Time ke har chote slice $dt$ mein, jo bhi abhi zinda hain unka wahi fraction marega. Us fraction-per-unit-time ko $\lambda$ (Greek letter "lambda") kahte hain. Toh ek slice $dt$ mein, deaths ki sankhya hai $\lambda \times (\text{those alive})\times dt$.

YEH assumption kyun, koi aur nahi. Real parts ki teen life stages hoti hain (dekho Bathtub Curve): early "infant" failures, ek lamba flat middle, aur budhape ki wear-out. Flat middle ke dauran death rate genuinely roughly constant hoti hai — ek working chip "as good as new" hota hai chahe woh kitne bhi time se chal raha ho. Woh memoryless middle precisely wahi hai jahan ek single number $\lambda$ kaam karta hai, aur spacecraft apna mission wahi spend karte hain.

Rate kyun, sirf count nahi. Ek raw death count depend karta hai bheed kitni badi hai. Ek rate — deaths per survivor per hour — part ki crowd-independent fingerprint hai.

PICTURE. Shrinking blue bar se har second same-size bites li jaati hain. Kyunki bar shrink karta hai, har equal-fraction bite pichli se kam actual dots remove karti hai — slowing decay ka seed yahi hai.

---

Step 3 — Picture ko change ke ek rule mein badalna

KYA. Chalte hain "survival fraction kitni tez girta hai" likhte hain. $R$ mein change ek tiny time $dt$ mein $\frac{dR}{dt}$ likha jaata hai — ise "blue curve ka slope" padho. Step 2 se, drop equal hai $\lambda$ times jo bhi abhi zinda hain, jo ki $R$ hi hai:

$\underset{\text{how fast }R\text{ falls}}{\underbrace{\frac{dR}{dt}}}=-\underset{\text{death rate}}{\underbrace{\lambda }}\underset{\text{who’s left to die}}{\underbrace{R}}$

Derivative tool kyun. Hum sirf nahi jaanna chahte kitne mare — hum curve ka instantaneous slope jaanna chahte hain, kyunki poora behaviour is baare mein hai ki drop kaise badalta rehta hai. "Abhi ka slope" precisely wahi hai jo symbol $\frac{dR}{dt}$ matlab hai: rise over run jab run zero ki taraf shrink ho.

Minus sign kyun. $R$ neeche jaata hai, toh iska slope negative hai. Minus sign yeh fact carry karta hai; iske saath, $\lambda$ ek clean positive number rehta hai.

PICTURE. Blue curve par har point par ek choti orange tangent arrow slope dikhati hai. Jahan curve zyada upar hai, arrow steep hai (bahut zinda → bahut mar rahe hain). Jahan curve neeche hai, arrow shallow hai (kam zinda → kam mar rahe hain). Slope height ke proportional hai. Woh ek sentence hi equation hai.

---

Step 4 — Solve karo: alag karo, phir slices add karo

KYA. Hum rearrange karte hain taaki saari $R$-wali cheezein ek taraf hon aur saari $t$-wali cheezein doosri taraf. Ise separating variables kehte hain:

$\underset{\text{tiny }\%\text{ drop in }R}{\underbrace{\frac{dR}{R}}}=-\lambda \underset{\text{tiny time step}}{\underbrace{dt}}$

Phir hum add up karte hain har tiny slice ko shuru se ($t=0$, $R=1$) ab tak ($t$, $R(t)$). "Infinitely many tiny slices add karo" exactly wahi hai jo integral sign $\int$ matlab hai:

${\int }_1^{R(t)}\frac{dR}{R}=-\lambda {\int }_0^{t}dt$

Pehle separate kyun. Hum ek waqt mein sirf ek variable integrate kar sakte hain. Har variable ko apni side par rakhne se hum each side ko independently total kar sakte hain — jaise folding se pehle laundry sort karna.

Logarithm kyun aata hai. Left side poochti hai: "kaunsa function, jab main iska slope leta hoon, $\frac{1}{R}$ deta hai?" Jawaab hai natural logarithm $\ln$. Toh $\ln$ hat se nahi nikala — yeh humpar force hota hai as the antidote to dividing by $R$. Right side easy hai: ek constant $\lambda$ ko time par add karo toh $\lambda t$ milta hai.

$\underset{\text{left side total}}{\underbrace{\ln R(t)-\ln 1}}=\underset{\text{right side total}}{\underbrace{-\lambda t}}$

Kyunki $\ln 1=0$ (kuch bhi zero power par 1 deta hai), yeh bas $\ln R(t)=-\lambda t$ hai.

PICTURE. $0$ se $t$ tak flat $-\lambda$ line ke neeche shaded area height $\lambda$ aur width $t$ ka ek rectangle hai — iska area $\lambda t$ hai. Integral literally woh grey area hai.

---

Step 5 — Log undo karo: $e^{-\lambda t}$ se milo

KYA. Hamare paas hai $\ln R(t)=-\lambda t$. $R$ ko free karne ke liye hum $\ln$ ka exact opposite apply karte hain: number $e$ ko ek power tak uthana. Kyunki $e^{\ln x}=x$ (woh cancel karte hain — yahi "opposite" ka matlab hai):

$R(t)=e^{-\lambda t}$

Number $e$ kyun. $e\approx 2.718$ woh ek base hai jiska khud ka growth rate khud ki value ke barabar hai — "slope proportional to height" ka unique fixed point. Yahi precisely woh property hai jo Step 3 ne maangi thi, toh $e$ convenience nahi hai; yeh woh ek matra base hai jo fit hota hai.

PICTURE. Yeh hai payoff curve. Yeh $R(0)=1$ se shuru hota hai, bahut tez girta hai jab bahut log zinda hote hain, phir flat ho jaata hai jab bheed thins hoti hai — exactly woh behaviour jo humne Step 3 mein sketch kiya tha, ab exact.

---

Step 6 — Edge aur degenerate cases (reader ko kabhi stranded mat chodo)

KYA & KYO. Ek model tabhi trustworthy hai jab woh apni extremes par sane behave kare. Hum sare char corners test karte hain.

Input	Formula deta hai	Matlab — kya yeh sense karta hai?
$t=0$	$R=e^0=1$	Launch par sab zinda. ✅
$t\to \infty$	$R=e^{-\infty }=0$	Forever diya, sab eventually mar jaate hain. ✅
$\lambda =0$	$R=e^0=1$ sab $t$ ke liye	Ek part jo fail nahi kar sakta forever jeeta hai. ✅
$t=1/\lambda$	$R=e^{-1}\approx 0.37$	"Average life" par sirf 37% abhi bhi zinda — half nahi!

PICTURE. Curve ko sare char checkpoints marked ke saath dikhaya gaya hai. $t=1/\lambda$ par green marker ko dekho: wahan height $0.37$ hai, $0.5$ nahi. Yeh poore topic ka sabse common misread hai.

---

Step 7 — Average lifetime nikaalo (MTTF)

KYA. Mean Time To Failure sabhi individual death-times ka average hai. Ek continuous spread ka average matlab hai: har possible death-time $t$ ko weight karo us instant par death kitni likely hai usse, phir add up karo. Woh weight hai $f(t)$, failure density, jo sirf $R$ ka downward slope hai:

$f(t)=-\frac{dR}{dt}=\lambda e^{-\lambda t}$

$\text{MTTF}={\int }_0^{\infty }\underset{\text{a death time}}{\underbrace{t}}\underset{\text{its likelihood}}{\underbrace{\lambda e^{-\lambda t}}}dt=\frac{1}{\lambda }$

Yeh $R(t)$ ke neeche area ke barabar kyun hai. Ek beautiful shortcut hai: kisi bhi cheez ke liye jo 1 se shuru ho aur sirf gire, average lifetime total area ke barabar hai survival curve ke neeche. Toh MTTF simply $e^{-\lambda t}$ ke neeche area hai, jo $1/\lambda$ hai.

PICTURE. Survival curve ke neeche shaded area ek "filled-in" exponential hai; iska area height 1 ke equal-area rectangle ki width ke barabar hai — woh width exactly $1/\lambda$ hai.

---

Ek-picture summary

Poora safar ek canvas par: parts ki bheed → constant death fraction $\lambda$ → slope = $-\lambda R$ → separate & integrate → $\ln$ aata hai → $e$ se undo karo → $R(t)=e^{-\lambda t}$ → neeche ka area = MTTF $=1/\lambda$.

Recall Feynman retelling — apne words mein wapas bolo

Ek hazaar identical gadgets line mein lagao aur on karo. Har second, jo bhi abhi bhi chal raha hai unka wahi slice quit karta hai — woh slice-per-second $\lambda$ hai. Kyunki bheed shrink karti hai, actually quit karne waalon ki sankhya bhi shrink hoti hai, toh survival curve pehle tezi se girta hai phir dheema pad jaata hai. "Drop equal hai $\lambda$ times jo bacha hai" likhna aur saare tiny drops add karna ek logarithm ko show up hone par force karta hai; log ko number $e$ se undo karne par $R(t)=e^{-\lambda t}$ milta hai. Corners test karo: launch par full, forever ke baad zero, aur — woh sneaky wala — "average" lifetime par sirf $37\%$ bacha, half nahi. Finally, average lifetime khud simply woh area hai jo us survival curve ke neeche hai, jo clean $1/\lambda$ nikalta hai. Yeh poora model hai, kuch memorise karne ki zaroorat nahi.

Recall Quick self-check

Constant failure rate hote hue bhi slowing death count kyun hota hai? ::: Kyunki constant ek fraction of survivors hai; jaise survivor pool shrink hota hai, wahi fraction kum aur kum actual parts remove karta hai. Derivation mein logarithm ko kyun force kiya jaata hai? ::: $\frac{dR}{R}$ integrate karna — log woh function hai jiska slope $1/R$ hai, toh yeh ek matra antidote hai. $t=1/\lambda$ par $R(t)$ kya hai? ::: $e^{-1}\approx 0.37$, toh mean lifetime tak $63\%$ already fail ho chuke hote hain. MTTF geometrically kya hai? ::: Survival curve $R(t)=e^{-\lambda t}$ ke neeche ka total area, jo $1/\lambda$ ke barabar hai.

---

Prerequisites & neighbours: Probability and Statistics Fundamentals · Poisson Process · Weibull Distribution · Bathtub Curve · Fault Tree Analysis · Availability vs. Reliability · Mission Design Constraints · Safety-Critical Systems