STATISTICAL METHODS FOR RELIABILITY

ANALYSIS

STAT4125A

Presented by Dr. Herbert Hove

INTRODUCTION:

Reliability: A measure of the ability of a technical

system/subsystem/component/service to perform its function when
required under specified conditions of use (environmental/operational)
over an intended period of time.

Reliability criteria:

1) Survival probability of a system: The probability that the

system does not fail within a given time interval

2) Availability of a system: The probability that the system is

working at a given time point (or able to do its job when needed)

3) Mean lifetime of a system: The expected time to the first system

failure

System failure:

The termination of the ability of the system to perform a required

function. A failure occurs when the function cannot be performed or
has a performance that falls outside the performance requirement.
Failure types

1) Sudden/catastrophic failures (bulb). Occur at any time, often

complete loss of function

2) Drift/degraded failures (tread of tyres). Drifts towards a failure

threshold, often performance fails to meet industrial standards
though they may still be residual functionality

Subjects of the mathematical theory of reliability:

1) Investigation of the mutual relationship between reliability

criteria of a system and reliability criteria of its subsystems
(elements, components)

2) Modeling of the failure and aging behavior of systems

3) Estimation of reliability criteria of systems

4) Maintenance modeling: Development, investigation and

optimization of measures for maintaining and restoring the
capability of a system to properly do its job
 Contents

1. Fundamental System Structures

2. Complex Systems
3. Monotone Systems with Dependent Lifetimes
4. Maintenance Theory
5. Reliability (Life) Data Analysis
 1 Fundamental System Structures

1.1 Simple Systems

A technical system is called simple if it can be considered a unit with

regard to all relevant reliability and maintenance aspects.

The basic reliability characteristic (criterion) of a technical (non-

repairable) system or of an organism is its (random – subject to chance
variations) lifetime.

The lifetime of a system is the time span from its starting up time point
(birth) to its (first) failure (death), where ‘failure’ is assumed to be an
instantaneous event (sudden failure only considered).

The term lifetime is used though time-to-failure will also be used (in
some cases). Unless stated otherwise, the system is assumed to be new
and in a functioning state at birth (t = 0).

Lifetime maybe continuous or discrete and may have units other than
“clock time”. ”.
Examples of more indirect time concepts include:

❖ Number of rotations of a bearing

❖ Number of times a switch is operated
❖ Number of kilometers driven by a car
❖ Number of cycles for a periodically working system

Observe from these examples that lifetime may often be a discrete

variable. A discrete variable can, however, be approximated by a
continuous variable.

Hence and unless stated otherwise, continuous lifetime is considered.

Let 𝑋 be a lifetime with distribution function

𝑥
𝐹 (𝑥) = 𝑃(𝑋 ≤ 𝑥) = ∫ 𝑓(𝑢)𝑑𝑢, 𝑥>0
0

The event {𝑇 ≤ 𝑡} occurs when the system fails before time 𝑡.

Hence, 𝐹 (𝑥) is called failure probability of a system with respect to an

interval [0, 𝑥], i.e. the probability that a system fails in [0, 𝑥].
The probability density function (pdf) f(𝑥) is defined as:

𝑑 𝐹 (𝑥 + ∆𝑥) − 𝐹 (𝑥)
𝑓 (𝑥 ) = 𝐹 (𝑥) = lim
𝑑𝑥 ∆𝑥→0 ∆𝑥

𝑃(𝑥 < 𝑋 ≤ 𝑥 + ∆𝑥)

= lim
∆𝑥→0 ∆𝑥

When ∆𝑥 is small,

𝑃(𝑥 < 𝑋 ≤ 𝑥 + ∆𝑥) ≈ 𝑓 (𝑥) ∙ ∆𝑥

where 𝑓 (𝑥) ∙ ∆𝑥 is the approximate probability that a system started at

birth will fail in the short interval (𝑥, 𝑥 + ∆𝑥].

The survival probability

𝐹̅ (𝑥) = 1 − 𝐹 (𝑥) = 𝑃(𝑋 > 𝑥) = 𝑅(𝑥), 𝑥>0

is the probability that the system does not fail in [0, 𝑥].

Equivalently,
𝑥 ∞
𝐹̅ (𝑥) = 1 − ∫ 𝑓 (𝑢)𝑑𝑢 = ∫ 𝑓(𝑢)𝑑𝑢
0 𝑥
The 1 − 𝛼 quantile of a pdf 𝑓(𝑥 ) is the lifetime value for which the
probability of non-exceedance is 1 − 𝛼.

That is,

𝑃(𝑋 ≤ 𝑞1−𝛼 ) = 1 − 𝛼
Residual (Remaining) Lifetime:

Assume the system has already worked for 𝑡 time units without failing
(𝑋 > 𝑡), and denote the residual lifetime 𝑋 − 𝑡 by 𝑋𝑡 . Then:

𝑃 (𝑡 < 𝑋 ≤ 𝑡 + 𝑥 )
𝐹𝑡 (𝑥) = 𝑃(𝑋𝑡 ≤ 𝑥) = 𝑃(𝑋 − 𝑡 ≤ 𝑥|𝑋 > 𝑡) =
𝑃 (𝑋 > 𝑡 )

Thus

𝐹 (𝑡 + 𝑥 ) − 𝐹 (𝑡 )
𝐹𝑡 (𝑥) = ; 𝑥 ≥ 0, 𝑡 ≥ 0
𝐹̅ (𝑡)

and

𝐹̅ (𝑡 + 𝑥 )
𝐹̅𝑡 (𝑥) = ; 𝑥 ≥ 0, 𝑡 ≥ 0
𝐹̅ (𝑡)

Expected residual lifetime 𝜇(𝑡) = 𝐸 (𝑋𝑡 ) of a system is:

∞
1
𝜇 (𝑡 ) = ∫ 𝐹̅ (𝑡 + 𝑥)𝑑𝑥
𝐹̅ (𝑡) 𝑡

Mean time to failure (𝜇 = 𝑀𝑇𝑇𝐹 < ∞) of a system is:

∞
𝜇 = 𝐸 (𝑋) = ∫ 𝐹̅ (𝑥) 𝑑𝑥
0
Example 1.1.1 (Uniform distribution): Let the random variable 𝑋 be
uniformly distributed over [0, 𝑇].

Then its distribution function is

0 𝑓𝑜𝑟 𝑥 < 0,
𝑥
𝐹 (𝑥 ) = { 𝑓𝑜𝑟 0 ≤ 𝑥 ≤ 𝑇,
𝑇
1 𝑓𝑜𝑟 𝑇 < 𝑥.

Hence the conditional failure probability is

𝐹 (𝑡 + 𝑥 ) − 𝐹 (𝑡 ) 𝑥
𝐹𝑡 (𝑥) = = ; 0 ≤ 𝑡 < 𝑇, 0≤𝑥 ≤𝑇−𝑡
𝐹̅ (𝑡) 𝑇−𝑡

Example 1.1.2 (Exponential distribution):

Let 𝑋 have an exponential distribution with parameter 𝜆, i.e.

𝑓 (𝑥) = 𝜆𝑒 −𝜆𝑥 ; 𝐹 (𝑥) = 1 − 𝑒 −𝜆𝑥 , 𝑥 > 0, 𝜆>0

Then:

1 − 𝑒 −𝜆(𝑡+𝑥) − (1 − 𝑒 −𝜆𝑡 )
𝐹𝑡 (𝑥) = −𝜆𝑡
= 1 − 𝑒 −𝜆𝑥
𝑒
Definition 1.1.1 (Aging system):

As system is AGING (rejuvenating) in [𝑡1 , 𝑡2 ], 𝑡1 < 𝑡2 if for any

arbitrary but fixed 𝑥, the conditional failure probability 𝐹𝑡 (𝑥) is
INCREASING (decreasing) in 𝑡, 𝑡1 ≤ 𝑡 ≤ 𝑡2 .

Provided the density function 𝑓 (𝑥 ) = 𝐹 ′ (𝑥) exists, the aging behaviour

of a system may be modelled based on the concept of its failure rate.

The probability that a system functioning at time 𝑡 will fail in the

interval (𝑡, 𝑡 + ∆𝑡] is:

𝐹 (𝑡 + ∆𝑡) − 𝐹 (𝑡)
𝐹𝑡 (∆𝑡) = 𝑃(𝑋𝑡 ≤ ∆𝑡) =
𝐹̅ (𝑡)

Dividing this conditional failure probability by the length of the time

interval, ∆𝑡.

This gives a conditional failure probability per unit time, i.e., a ‘failure
probability rate’ with units “per time”:

1 𝐹 (𝑡 + ∆𝑡) − 𝐹 (𝑡) 1
𝐹𝑡 (∆𝑡) = ∙
∆𝑡 ∆𝑡 𝐹̅ (𝑡)
For ∆𝑡 → 0, the first ratio on the right-hand side tends to 𝑓 (𝑡). Hence,

1 𝑓 (𝑡 )
lim ( )
𝐹𝑡 ∆𝑡 =
∆𝑡→0 ∆𝑡 𝐹̅ (𝑡)

This limit is called failure rate or hazard function and is denoted as

𝜆(𝑡):

𝑓 (𝑡 )
𝜆 (𝑡 ) =
𝐹̅ (𝑡)

𝜆(𝑡) is often more convenient to work with than the density function
𝑓 (𝑡 ).

For example, suppose that 𝑓 (𝑡) is exponential, that is:

𝑓(𝑡) = 𝜆𝑒 −𝜆𝑡 ; 𝐹 (𝑡) = 1 − 𝑒 −𝜆𝑡 ; 1 − 𝐹 (𝑡) = 𝑒 −𝜆𝑡 .

Hence,

𝜆𝑒 −𝜆𝑡
𝜆(𝑡) = −𝜆𝑡 = 𝜆
𝑒

Thus, an exponential failure density corresponds to a constant failure

rate with respect to time.
Integrated failure rate:

𝑑
Since 𝑓 (𝑡) = 𝐹 (𝑡) = −𝑅′ (𝑡),
𝑑𝑡

then

−𝑅′ (𝑡) 𝑑
𝜆 (𝑡 ) = = − ln𝑅(𝑡).
𝑅 (𝑡 ) 𝑑𝑡

Hence,
𝑥 𝑥
− ∫0 𝜆(𝑡)𝑑𝑡
∫ 𝜆(𝑡)𝑑𝑡 = −ln𝑅(𝑥) ⟺ 𝑅(𝑥) = 𝑒 , 𝑥≥0
0

Introducing the integrated failure rate:

𝑥
Λ(𝑥) = ∫ 𝜆(𝑡)𝑑𝑡
0

Then 𝐹 (𝑥), 𝐹𝑡 (𝑥) and the corresponding survival probabilities for

𝑥 ≥ 0 and 𝑡 ≥ 0 can be written as:

𝐹 (𝑥) = 1 − 𝑒 −Λ(𝑥)

𝐹𝑡 (𝑥) = 1 − 𝑒 −[Λ(𝑡+𝑥)−Λ(𝑡)]

𝐹̅ (𝑥) = 𝑒 −Λ(𝑥)

𝐹̅𝑡 (𝑥) = 𝑒 −[Λ(𝑡+𝑥)−Λ(𝑡)]

Thus if 𝑋 has a probability density, then its probability distribution is
fully characterised by its failure rate and vice versa.

The above representation of 𝐹̅𝑡 (𝑥) implies an important property of the

failure rate:

A system AGES (rejuvenates) in [𝑡1 , 𝑡2 ], 𝑡1 < 𝑡2 , if its failure rate 𝜆(𝑡)

is INCREASING (decreasing) in [𝑡1 , 𝑡2 ].

The failure rate curve is usually called a bathtub curve after its
characteristic shape.
Example 1.1.3 (Weibull distribution): Let 𝑋 has a Weibull
distribution with distribution function

𝑥 𝛽
−( )
𝐹 (𝑥 ) = 1 − 𝑒 𝜃 ,𝑥 ≥ 0; 𝜃 > 0, 𝛽 > 0

where 𝛽 is the shape parameter and 𝜃 is the scale parameter.

Then:

𝛽 𝑥 𝛽−1 −(𝑥 )𝛽
𝑓 (𝑥 ) = ( ) 𝑒 𝜃 ,𝑥 ≥ 0
𝜃 𝜃

𝛽 𝑥 𝛽−1
𝜆 (𝑥 ) = ( ) ,𝑥 > 0
𝜃 𝜃

If

𝛽 > 1, 𝜆(𝑥) it is increasing in [0, ∞)

𝛽 < 1, 𝜆(𝑥) it is decreasing in [0, ∞)

𝛽 = 1, 𝜆(𝑥) is identically constant: 𝜆(𝑡) ≡ 𝜆 = 1⁄𝜃 (i.e. 𝑋 has an

exponential distribution)
 1.2 Nonparametric Classes of Probability Distributions

Obvious classes: Increasing failure rate (IFR) and decreasing failure

rate (DFR).

IFR (DFR) Distribution: 𝐹 (𝑥) is an IFR (DFR) distribution if 𝐹𝑡 (𝑥)

is increasing (decreasing) in 𝑡 for fixed but arbitrary 𝑥.

If the density 𝑓(𝑥) = 𝐹 ′ (𝑥) exists, then 𝐹 (𝑥) is IFR (DFR) if and only
if the corresponding failure rate 𝜆(𝑥) increases (decreases) as a
function of 𝑥, for 𝑥 > 0.

Slightly more generally, 𝐹 (𝑥) is an IFR (DFR) distribution if −ln𝑅(𝑥)

is a convex (concave) function of 𝑥.

This is so since a differentiable convex (concave) function has an

increasing (decreasing) derivative.

In the following examples, some common lifetime distributions are

checked whether they are IFR, DFR or neither.
Example 1.2.1 (Uniform distribution): Let the random variable 𝑋 be
uniformly distributed over (0, 𝑇]. Then:

𝑥
𝐹 (𝑥 ) = , 0<𝑥≤𝑇
𝑇

1
𝑓 (𝑥 ) = , 0<𝑥≤𝑇
𝑇

Thus,

1⁄𝑇 1
𝜆 (𝑥 ) = = , 0<𝑥≤𝑇
1 − 𝑥 ⁄𝑇 𝑇 − 𝑥

which is strictly increasing for 0 < 𝑥 ≤ 𝑇. Accordingly, the uniform

distribution is IFR.

Example 1.2.2 (Exponential distribution): Let 𝑋 be exponentially

distributed with probability density:

𝑓(𝑥) = 𝜆𝑒 −𝜆𝑥 , 𝑥 > 0, 𝜆 > 0.

Hence,

𝜆(𝑥) = 𝜆, 𝑥>0

Describe the behaviour of the Exponential distribution failure rate, and

hence, its nonparametric class.
Example 1.2.3 (Weibull distribution):

Let 𝑋 be Weibull distributed with parameters 𝛽 > 0 and 𝜃 > 0 having

distribution function:

𝑥 𝛽
−( )
𝐹 (𝑥 ) = 1 − 𝑒 𝜃 , 𝑥 > 0, 𝛽 > 0, 𝜃 > 0.

It follows that

𝛽 𝑥 𝛽−1
𝜆 (𝑥 ) = ( ) , 𝑥 > 0.
𝜃 𝜃

If:

❖ 𝛽 > 1, 𝐹 (𝑥) is IFR

❖ 𝛽 < 1, 𝐹 (𝑥) is DFR

❖ 𝛽 = 1 𝐹 (𝑥) is both IFR and DFR (exponential distribution)

Technical systems have periods in their life when they do not age.
Examples include:

❖ periods of no or reduced stress

❖ when maintenance actions such as repairs rejuvenate the system

To model these more general concepts of aging and rejuvenation,

distribution functions of type increasing failure rate average (IFRA)
or decreasing failure rate average (DFRA) are more suitable.

IFRA (DFRA) Distribution: 𝐹 (𝑥) is an IFRA (DFRA) distribution if

1 𝑡 1 1
𝜆̅(𝑡) = ∫ 𝜆(𝑥)𝑑𝑥 = Λ(𝑡) = − ln𝐹̅ (𝑡)
𝑡 0 𝑡 𝑡

is increasing (decreasing) in 𝑡 in the interval [0, 𝑡].