Probability & Statistics

L EC T U R E 4 : C O N T INU OUS PR O BABIL IT Y D IST R IBU T IONS

D r Pe te r D . D u n n i n g

ENGINEERING RISK & RELIABILITY ANALYSIS

Normal distribution
o Most important continuous probability distribution in statistics
o Completely described by first two moments, the mean and variance, has the
characteristic symmetric bell‐shaped curve
o Good approximation that describes observations in many phenomena that
occur in nature, industry and research
o Physical measurements: e.g. weather data, sizes of manufactured parts
o Errors in scientific measurements: i.e. accuracy of measuring method

o Discovery of the mathematical equation usually attributed to Karl Freidrich

Gauss, who derived the equation from a study of errors in repeated
measurements of the same quantity, published in1809
Carl Friedrich Gauss
o Thus, the normal distribution is also called the Gaussian distribution (1777–1855)

ENGINEERING RISK & RELIABILITY ANALYSIS 2

Normal distribution
o The normal distribution describes the PDF of a normal random variable X
with mean value μx and variance 𝜎𝑥2

2
1 1 𝑥 − 𝜇𝑥
PDF: 𝑓 𝑥 = exp −
𝜎𝑥 2𝜋 2 𝜎𝑥

𝑧
1 2
CDF: 𝑃 𝑋≤𝑎 =𝐹 𝑎 = න 𝑒 −0.5𝑣 𝑑𝑣
2𝜋
−∞

Standard normal variable: 𝑧 = 𝑎 − 𝜇𝑥 /𝜎𝑥

ENGINEERING RISK & RELIABILITY ANALYSIS 3

Normal distribution
o The evaluation of the normal CDF integral is not straightforward
o Therefore, tabulated (pre-calculated) values are usually used
o However, it is difficult to setup separate tables for every conceivable mean and
standard deviation value
o All normal distributions can be transformed to the standard normal distribution with
mean value 0 and variance 1 using the standard normal variable z :
𝑧 𝑥 = 𝑥 − 𝜇𝑥 /𝜎𝑥 z (x) = number of standard deviations from the mean
Then: 1
PDF: 𝑓 𝑥 = 𝜙 𝑧 𝑥
𝜎𝑥
CDF: 𝑃 𝑋≤𝑎 =𝐹 𝑎 =Φ 𝑧 𝑎

ENGINEERING RISK & RELIABILITY ANALYSIS 4

Normal distribution
o Tables of standard normal CDF values often start at 0 and include values for z ≤ 0

o Values for positive z can be obtained by using: Φ 𝑧 = 1 − Φ −𝑧

o The tables can also be used to evaluate the inverse CDF function: 𝑧 = Φ−1 𝑝

𝑃 𝑍 ≤ 0.35 = Φ 0.35
= 1 − Φ −0.35 = 𝟎. 𝟔𝟑𝟔𝟖𝟑

𝑧 = Φ−1 0.2 ≈ −𝟎. 𝟖𝟒

ENGINEERING RISK & RELIABILITY ANALYSIS 5

Properties of the normal distribution
o Symmetry: the probability that X is greater than a:
𝑃 𝑋 > 𝑎 = Φ −𝑧 𝑎 =1−Φ 𝑧 𝑎

o Inverse: given a probability of X less than or equal to a, the inverse CDF

function can be used to find the value of z :
𝑧 = Φ−1 𝑝 = −Φ−1 1 − 𝑝 , 𝑝 = 𝑃(𝑋 ≤ 𝑎)
o Probability that X is between a and b is the difference of the CDF values:
𝑃 𝑎 <𝑋 ≤𝑏 =𝐹 𝑏 −𝐹 𝑎 =Φ 𝑧 𝑏 −Φ 𝑧 𝑎
a b
o Summation: if 𝑌 = σ 𝑋𝑖 and each variable X is normal and independent then 𝑌
is a normal random variable with:
𝜇𝑌 = σ 𝜇𝑋 and 𝜎𝑌2 = σ 𝜎𝑋2

ENGINEERING RISK & RELIABILITY ANALYSIS 6

Example
o The measurement of temperature in a
pipeline follows a normal distribution with
mean of 350°K and variance of 100 (°K2).

1. What is the probability the measured temperature exceeds 365°K ?

First, calculate standard normal variable: z = (x – μ)/σ = (365 – 350)/10 = 1.5

Then: P(X > 365) = P(Z > 1.5) = 1 – P(Z ≤ 1.5)

= 1 – Φ(1.5) = 1 – 1 + Φ(-1.5) = 0.06681

ENGINEERING RISK & RELIABILITY ANALYSIS 7

Example
2. What is the probability the measured temperature is between 335
and 360°K ?
Calculate the standard normal variables:
za = (335 – 350)/10 = -1.5 zb = (360 – 350)/10 = 1
Then: P(335 < X ≤ 350) = P(-1.5 < Z ≤ 1) = Φ(1.0) - Φ(-1.5)
= 1 - Φ(-1.0) - Φ(-1.5) = 1 – 0.15866 – 0.06681
= 0.77453
3. Determine the temperature for which the probability of the
measurement being below this value is 0.98.
First use: z = -Φ-1(1 – p) = -Φ-1(0.02) ≈ 2.055
Then: x = z×σ + μ = 2.055×10 + 350 = 370.55°K

ENGINEERING RISK & RELIABILITY ANALYSIS 8

Example
o The total length of a component is measured in two parts. It is known that
these measurements follow a normal distribution. The mean measurement
values are 150 mm and 120 mm, and standard deviations 2.0 and 1.5.
Assume the measurements are independent.
1. What is the probability that the component is longer than 273 mm?
First, the total length is also normally distributed with:
μ = 150 + 120 = 270 mm , σ = (22 + 1.52)0.5 = 2.5
Now, transform to a standard normal variable:
z = (273 – 270) / 2.5 = 1.2
Finally, evaluate the normal CDF using the table:
P(X > 273) = P(Z > 1.2) = 1 – [1 - Φ(-1.2)] = 0.11507

ENGINEERING RISK & RELIABILITY ANALYSIS 9

Central limit theorem
o Under some (common) conditions, the probability distribution for the sum of a large
number of statistically independent random variables approaches the normal distribution,
irrespective of the individual distributions of the random variables
o The central limit theorem can be used to approximate other distributions under certain
conditions:
➢ Binomial distribution: if n (number of trials) is large and p not close to zero or one, then the
distribution is approximately normal with mean np and variance np(1−p)

➢ Poisson distribution: if the Poisson parameter, λ is large, then the distribution is approximately
normal with mean λ and variance λ

ENGINEERING RISK & RELIABILITY ANALYSIS 10

Lognormal distribution
o If a continuous random variable X is the product of independent random variables:
X = X1 X2 X3 … Xn
o Then, the probability distribution of X tends to be lognormal
o The name of the distribution comes from the fact that if X is lognormal, then the
distribution of ln X is normal. This is expected as:
ln X = ln X1 + ln X2 + ln X3 + … + ln Xn
o The values of X have to be strictly positive (X > 0)
o Lognormal distribution is very useful when the values of the random variable are
known, from physical consideration, to be strictly positive
o Example: strength, stiffness and fatigue life of materials

ENGINEERING RISK & RELIABILITY ANALYSIS 11

Lognormal distribution
o If a random variable x is lognormal, then:
2
1 1 ln 𝑥 − 𝜆
PDF: 𝑓 𝑥 = exp − ,0 ≤ 𝑥
𝑥𝜁 2𝜋 2 𝜁

Parameters: 𝜆 = 𝐸 ln 𝑥 𝜁 2 = var ln 𝑥 λ=0

o The value of they CDF can be determined from the
normal CDF function table:
ln 𝑎 − 𝜆
CDF: 𝑃 𝑋≤𝑎 =𝐹 𝑎 =Φ
𝜁

Mean: 𝜇𝑥 = exp 𝜆 + 0.5𝜁 2

Variance: 𝜎𝑥2 = 𝜇𝑥2 exp 𝜁 2 − 1

ENGINEERING RISK & RELIABILITY ANALYSIS 12

Lognormal distribution
o Properties: 2
𝜎𝑥
Parameter relations: 𝜆 = ln 𝜇𝑥 − 0.5𝜁 2 𝜁 2 = ln 1 + = ln 1 + 𝑉𝑥2
𝜇𝑥

Median: 𝑃 𝑋 ≤ 𝑥𝑚 = 0.5 → 𝐹 𝑥𝑚 = 0.5 → 𝜆 = ln 𝑥𝑚

ln 𝜇𝑥
𝑥𝑚 = 𝑒𝜆 =
exp 0.5𝜁 2

oThe normal and lognormal distributions are related

o Thus, the central limit theorem can be extended for the lognormal distribution
o The distribution of the product of a large number of uncertain variables approaches
the lognormal distribution, irrespective of the individual distributions

ENGINEERING RISK & RELIABILITY ANALYSIS 13

Example
o The yield strength of a type of steel follows a lognormal distribution with
a mean of 228 MPa and a coefficient of variance 0.12.
o It is proposed to use the steel in an structure that will experience a
maximum stress of 170 MPa.
1. What is the probability that the structure will yield?
First, calculate parameters of the lognormal distribution:
𝜁 2 = ln 1 + 0.12 2
= 0.0143 , 𝜻 = 𝟎. 𝟏𝟐𝟎 𝜆 = ln 228 − 0.5 × 0.0143 = 𝟓. 𝟒𝟐𝟐

The probability is then: 𝑃 𝑋 ≤ 170 = 𝐹 170

ln 170 − 5.422
=Φ = Φ −2.385 = 𝟎. 𝟎𝟎𝟖𝟓
0.12

ENGINEERING RISK & RELIABILITY ANALYSIS 14

Exponential distribution
o The recurrence time for a Bernoulli process is described by the geometric distribution
(as seen in lecture 3)
o For a Poisson process, the recurrence time is described by the exponential distribution
o The Poisson distribution describes the probability of an event occurring a number of
times in a particular period of time or region of space
o In many applications, the time period or spatial region is the random variable
o Thus, can be modelled using an exponential distribution

ENGINEERING RISK & RELIABILITY ANALYSIS 15

Exponential distribution
o Random variable: time (or spatial dimension) to first occurrence of an event
o The event occurs according to a Poisson process
o Exponential distribution: probability that x is the time (or spatial dimension) to first
occurrence of an event, with mean rate of occurrence v:
PDF: 𝑓 𝑥 = 𝑣𝑒 −𝑣𝑥 , 0 ≤ 𝑥
𝑎

CDF: 𝑃 𝑋 ≤ 𝑎 = 𝐹 𝑎 = න 𝑣𝑒 −𝑣𝑥 𝑑𝑥 = 1 − 𝑒 −𝑣𝑎

0

Mean: 𝜇𝑥 = 1/𝑣 Average time between events

Variance: 𝜎𝑥2 = 1/𝑣 2

ENGINEERING RISK & RELIABILITY ANALYSIS 16

Exponential distribution
o The exponential distribution has no memory, it is the only continuous
distribution taking positive values with this property:
Event does not occur before t0, probability event
𝑃 𝑋 ≥ 𝑡0 + 𝑡 | 𝑋 ≥ 𝑡0 = 𝑃 𝑋 ≥ 𝑡 occurs at t0 + t is the same as if it occurring at time t

o Memoryless property important to consider if using an exponential distribution

to calculate the reliability of components or machines in lifetime problems
o Example: using an exponential distribution, a component that does not fail up
to 1,000 hours, has the same probability of lasting an additional 500 hours as
the probability of lasting the first 500 hours
o Thus, there is no consideration of time factors, such as wear, that may
influence the reliability of a component, given it has operated for t0 hours

ENGINEERING RISK & RELIABILITY ANALYSIS 17

Example
o A factory produces steel bars and the probability that a defect is
detected in a length of bar can be modelled as a Poisson process,
with mean rate of occurrence of 1 defect in 50 meters of bar.

1. What is the probability that a 10 m bar will have no detectable defects?

v = 1/50 = 0.02 , P(X > 10) = 1 – P(X ≤ 10) = 1 – [1 – e-0.02×10] = 0.819
2. What is the probability that the next defect is detected between 5 and 8 m?
P(5 < X < 8) = F(8) – F(5) = e-0.02×5 – e-0.02×8 = 0.053

ENGINEERING RISK & RELIABILITY ANALYSIS 18

Shifted exponential distribution
o The exponential distribution starts at X = 0, but we can start it at any point X > 0. This is
called the shifted exponential distribution.
o If the a shifted exponential distribution starts at Y > 0 then:

PDF: 𝑓 𝑥 = 𝑣𝑒 −𝑣 𝑥−𝑌
,𝑌 ≤ 𝑥 < ∞

CDF: 𝐹 𝑎 = 1 − 𝑒 −𝑣 𝑎−𝑌

Mean: 𝜇𝑥 = 𝑌 + 1/𝑣

Variance: 𝜎𝑥2 = 1/𝑣 2

ENGINEERING RISK & RELIABILITY ANALYSIS 19

Weibull distribution
o The Weibull (vie-bool) distribution is often used to model time to failure
o The parameters in the distribution provide flexibility to model systems
in which the number of failures:
➢ increases with time (wear out effect),

➢ decreases with time (wear in effect) or

➢ remains constant with time (failures caused by external shocks to system)

o Named after Waloddi Weibull, a Swedish engineer, scientist, and

mathematician who used the distribution in reliability analysis Waloddi Weibull
(1887 - 1979),

ENGINEERING RISK & RELIABILITY ANALYSIS 20

Weibull distribution
o The Weibull distribution has two parameters:
➢ Scale parameter: k
β = 1 , k = 0.5
➢ Shape parameter: β f(x) β = 2 , k = 0.5

o The distribution is then defined as: β = 3 , k = 0.5

𝛽 𝑥 𝛽−1 𝑥 𝛽
PDF: 𝑓 𝑥 = exp − ,0 ≤ 𝑥
𝑘 𝑘 𝑘
x
𝛽
CDF: 𝑃 𝑋 ≤ 𝑎 = 𝐹 𝑎 = 1 − exp − 𝑎Τ𝑘

o The Weibull distribution is related to several other distributions:

o If β = 1, then it becomes the exponential distribution with mean rate of occurrence v = 1/k

ENGINEERING RISK & RELIABILITY ANALYSIS 21

Example
o Assume the life of a component can be modelled with a Weibull distribution with
β = 3 and k = 10,000 hours.
1. What is the probability that the component will last less than 10,000 hours?
𝑃 𝑋 ≤ 10,000 = 𝐹 10,000 = 1 − exp − 10000Τ10000 3

= 1 − exp −1 = 𝟎. 𝟔𝟑𝟐𝟏
2. What is the probability that the component will last longer than 5,000 hours?
𝑃 𝑋 > 5,000 = 1 − 𝑃 𝑋 ≤ 5,000 = 1 − 1 − exp − 5000Τ10000 3

= exp −0.53 = 𝟎. 𝟖𝟖𝟐𝟓

ENGINEERING RISK & RELIABILITY ANALYSIS 22

Summary of continuous distributions
Distribution PDF, f(x) CDF, P(X ≤ a) = F(a)
2
1 1 𝑥 − 𝜇𝑥 𝑎 − 𝜇𝑥
Normal 𝜎𝑥 2𝜋
exp −
2 𝜎𝑥
Φ
𝜎𝑥

2
1 1 ln 𝑥 − 𝜆 ln 𝑎 − 𝜆
Lognormal exp − ,0 ≤ 𝑥 Φ
𝑥𝜁 2𝜋 2 𝜁 𝜁

Exponential 𝑣𝑒 −𝑣𝑥 , 0 ≤ 𝑥 1 − 𝑒 −𝑣𝑎

𝛽 𝑥 𝛽−1 𝑥 𝛽
𝛽
Weibull 𝑘 𝑘
exp −
𝑘
,0 ≤ 𝑥 1 − exp − 𝑎Τ𝑘

ENGINEERING RISK & RELIABILITY ANALYSIS 23