c1

2
System Reliability Concepts
The analysis of the reliability of a system must be based on precisely defined

concepts. Since it is readily accepted that a population of supposedly identical
systems, operating under similar conditions, fall at different points in time, then a
failure phenomenon can only be described in probabilistic terms. Thus, the
fundamental definitions of reliability must depend on concepts from probability
theory. This chapter describes the concepts of system reliability engineering. These
concepts provide the basis for quantifying the reliability of a system. They allow
precise comparisons between systems or provide a logical basis for improvement
in a failure rate. Various examples reinforce the definitions as presented in Section
2.1. Section 2.2 examines common distribution functions useful in reliability
engineering. Several distribution models are discussed and the resulting hazard
functions are derived. Section 2.3 describes a new concept of systemability.
Several systemability functions of various system configurations such as series,
parallel, and k-out-of-n, are presented. Section 2.4 discusses various reliability
aspects of systems with multiple failure modes. Stochastic processes including
Markov process, Poisson process, renewal process, quasi-renewal process, and
nonhomogeneous Poisson process are discussed in Sections 2.5 and 2.6.
In general, a system may be required to perform various functions, each of
which may have a different reliability. In addition, at different times, the system
may have a different probability of successfully performing the required function
under stated conditions. The term failure means that the system is not capable of
performing a function when required. The term capable used here is to define if the
system is capable of performing the required function. However, the term capable
is unclear and only various degrees of capability can be defined.
10 System Software Reliability
2.1 Reliability Measures

The reliability definitions given in the literature vary between different practition-
ers as well as researchers. The generally accepted definition is as follows.
Definition 2.1: Reliability is the probability of success or the probability that the
system will perform its intended function under specified design limits.
More specific, reliability is the probability that a product or part will operate
properly for a specified period of time (design life) under the design operating
conditions (such as temperature, volt, etc.) without failure. In other words,
reliability may be used as a measure of the system’s success in providing its
function properly. Reliability is one of the quality characteristics that consumers
require from the manufacturer of products.
Mathematically, reliability R(t) is the probability that a system will be
successful in the interval from time 0 to time t:
R(t ) = P(T > t ) t≥0 (2.1)
where T is a random variable denoting the time-to-failure or failure time.
Unreliability F(t), a measure of failure, is defined as the probability that the sys-
tem will fail by time t:
F (t ) = P (T ≤ t ) for t ≥ 0
In other words, F(t) is the failure distribution function. If the time-to-failure
random variable T has a density function f(t), then
∞
R (t ) = ∫ f ( s)ds
t
or, equivalently,
d
f (t ) = − [ R (t )]
dt
The density function can be mathematically described in terms of T:
lim P(t < T ≤ t + Δt )
Δt → 0
This can be interpreted as the probability that the failure time T will occur between
the operating time t and the next interval of operation, t+Δt.
Consider a new and successfully tested system that operates well when put
into service at time t = 0. The system becomes less likely to remain successful as
the time interval increases. The probability of success for an infinite time interval,
of course, is zero.
Thus, the system functions at a probability of one and eventually decreases to
a probability of zero. Clearly, reliability is a function of mission time. For example,
one can say that the reliability of the system is 0.995 for a mission time of 24
hours. However, a statement such as the reliability of the system is 0.995 is
meaningless because the time interval is unknown.
System Reliability Concepts 11
Example 2.1: A computer system has an exponential failure time density function
t
1 −
f (t ) = e 9,000 t≥0
9, 000
What is the probability that the system will fail after the warranty (six months or
4380 hours) and before the end of the first year (one year or 8760 hours)?
Solution: From equation (2.1) we obtain

8760 t
1 −
P(4380 < T ≤ 8760) = ∫
4380
9000
e 9000 dt
= 0.237
This indicates that the probability of failure during the interval from six months to
one year is 23.7%.
If the time to failure is described by an exponential failure time density
function, then
t
1 −
f (t ) = e θ t ≥ 0, θ > 0 (2.2)
θ
and this will lead to the reliability function
∞ s t
1 − −
R (t ) = ∫θ t
e θ ds = e θ t≥0 (2.3)
Consider the Weibull distribution where the failure time density function is
given by
⎛ t ⎞β
β t β −1 −⎜⎜ θ ⎟⎠
f (t ) = β e ⎝ t ≥ 0, θ > 0, β > 0
θ
Then the reliability function is
⎛ t ⎞β
−⎜ ⎟
⎜
⎝θ ⎠
R (t ) = e t≥0
Thus, given a particular failure time density function or failure time distribution
function, the reliability function can be obtained directly. Section 2.2 provides
further insight for specific distributions.
System Mean Time to Failure

Suppose that the reliability function for a system is given by R(t). The expected
failure time during which a component is expected to perform successfully, or the
system mean time to failure (MTTF), is given by
∞
MTTF = ∫ t f (t )dt (2.4)
0
Substituting
d
f (t ) = −
[ R (t )]
dt
into equation (2.4) and performing integration by parts, we obtain
∫
MTTF = − td [ R (t )]
0
∞
(2.5)
∞
0
∫
= [−tR(t )] | + R(t )dt
0
The first term on the right-hand side of equation (2.5) equals zero at both
limits, since the system must fail after a finite amount of operating time. Therefore,
we must have tR(t) → 0 as t → ∞ . This leaves the second term, which equals
∞
∫
MTTF = R (t )dt
0
(2.6)
Thus, MTTF is the definite integral evaluation of the reliability function. In

general, if λ (t ) is defined as the failure rate function, then, by definition, MTTF is
not equal to 1/ λ (t ) .
The MTTF should be used when the failure time distribution function is
specified because the reliability level implied by the MTTF depends on the
underlying failure time distribution. Although the MTTF measure is one of the
most widely used reliability calculations, it is also one of the most misused
calculations. It has been misinterpreted as “guaranteed minimum lifetime”.
Consider the results as given in Table 2.1 for a twelve-component life duration test.
Table 2.1. Results of a twelve-component life duration test
Component Time to failure

(hours)
1 4510
2 3690
3 3550
4 5280
5 2595
6 3690
7 920
8 3890
9 4320
10 4770
11 3955
12 2750
Using a basic averaging technique, the component MTTF of 3660 hours was
estimated. However, one of the components failed after 920 hours. Therefore, it is
important to note that the system MTTF denotes the average time to failure. It is
neither the failure time that could be expected 50% of the time, nor is it the guaran-
teed minimum time of system failure.
A careful examination of equation (2.6) will show that two failure distributions
can have the same MTTF and yet produce different reliability levels. This is
illustrated in a case where the MTTFs are equal, but with normal and exponential
failure distributions. The normal failure distribution is symmetrical about its mean,
thus
R( MTTF ) = P ( Z ≥ 0) = 0.5
where Z is a standard normal random variable. When we compute for the
exponential failure distribution using equation (2.3), recognizing that θ = MTTF,
the reliability at the MTTF is
MTTF
−
R( MTTF ) = e MTTF = 0.368
Clearly, the reliability in the case of the exponential distribution is about 74% of
that for the normal failure distribution with the same MTTF.
Failure Rate Function

The probability of a system failure in a given time interval [t1, t2] can be expressed
in terms of the reliability function as
t2 ∞ ∞
∫
t1
∫ ∫
f (t )dt = f (t )dt − f (t )dt
t1 t2
= R(t1 ) − R(t2 )
or in terms of the failure distribution function (or the unreliability function) as
t2 t2 t1
∫ f (t )dt = ∫
t1 −∞
f (t )dt − ∫
−∞
f (t )dt
= F (t2 ) − F (t1 )
The rate at which failures occur in a certain time interval [t1, t2] is called the
failure rate. It is defined as the probability that a failure per unit time occurs in the
interval, given that a failure has not occurred prior to t1, the beginning of the
interval. Thus, the failure rate is
R (t1 ) − R (t2 )
(t2 − t1 ) R(t1 )
Note that the failure rate is a function of time. If we redefine the interval as
[t , t + Δt ], the above expression becomes
R (t ) − R(t + Δt )
ΔtR(t )
The rate in the above definitions is expressed as failures per unit time, when in
reality, the time units might be in terms of miles, hours, etc. The hazard function is
defined as the limit of the failure rate as the interval approaches zero. Thus, the
hazard function h(t) is the instantaneous failure rate, and is defined by
R (t ) − R(t + Δt )
h(t ) = lim
Δt → 0 ΔtR(t )
1 d
= [− R(t )] (2.7)
R(t ) dt
f (t )
=
R (t )
The quantity h(t)dt represents the probability that a device of age t will fail in the
small interval of time t to (t+dt). The importance of the hazard function is that it
indicates the change in the failure rate over the life of a population of components
by plotting their hazard functions on a single axis. For example, two designs may
provide the same reliability at a specific point in time, but the failure rates up to
this point in time can differ.
The death rate, in statistical theory, is analogous to the failure rate as the force
of mortality is analogous to the hazard function. Therefore, the hazard function or
hazard rate or failure rate function is the ratio of the probability density function
(pdf) to the reliability function.
Maintainability
When a system fails to perform satisfactorily, repair is normally carried out to
locate and correct the fault. The system is restored to operational effectiveness by
making an adjustment or by replacing a component.
Maintainability is defined as the probability that a failed system will be
restored to specified conditions within a given period of time when maintenance is
performed according to prescribed procedures and resources. In other words,
maintainability is the probability of isolating and repairing a fault in a system
within a given time. Maintainability engineers must work with system designers to
ensure that the system product can be maintained by the customer efficiently and
cost effectively. This function requires the analysis of part removal, replacement,
tear-down, and build-up of the product in order to determine the required time to
carry out the operation, the necessary skill, the type of support equipment and the
documentation.
Let T denote the random variable of the time to repair or the total downtime. If
the repair time T has a repair time density function g(t), then the maintainability,
V(t), is defined as the probability that the failed system will be back in service by
time t, i.e.,
t
V (t ) = P(T ≤ t ) = ∫ g ( s )ds
0
For example, if g (t ) = μ e− μ t where μ > 0 is a constant repair rate, then
V (t ) = 1 − e − μ t
which represents the exponential form of the maintainability function.
An important measure often used in maintenance studies is the mean time to
repair (MTTR) or the mean downtime. MTTR is the expected value of the random
variable repair time, not failure time, and is given by
∞
MTTR = ∫ tg (t )dt
0
When the distribution has a repair time density given by g (t ) = μ e− μ t , then, from
the above equation, MTTR = 1/ μ . When the repair time T has the log normal
density function g(t), and the density function is given by
(ln t − μ )2
1 −
g (t ) = e 2σ 2
t>0
2π σ t
then it can be shown that
σ2
MTTR = m e 2
where m denotes the median of the log normal distribution.
In order to design and manufacture a maintainable system, it is necessary to
predict the MTTR for various fault conditions that could occur in the system. This
is generally based on past experiences of designers and the expertise available to
handle repair work.
The system repair time consists of two separate intervals: passive repair time
and active repair time. Passive repair time is mainly determined by the time taken
by service engineers to travel to the customer site. In many cases, the cost of travel
time exceeds the cost of the actual repair. Active repair time is directly affected by
the system design and is listed as follows:
1. The time between the occurrence of a failure and the system user
becoming aware that it has occurred.
2. The time needed to detect a fault and isolate the replaceable
component(s).
3. The time needed to replace the faulty component(s).
4. The time needed to verify that the fault has been corrected and the system
is fully operational.
The active repair time can be improved significantly by designing the system in
such a way that faults may be quickly detected and isolated. As more complex
systems are designed, it becomes more difficult to isolate the faults.
Availability
Reliability is a measure that requires system success for an entire mission time. No
failures or repairs are allowed. Space missions and aircraft flights are examples of
systems where failures or repairs are not allowed. Availability is a measure that
allows for a system to repair when failure occurs.
The availability of a system is defined as the probability that the system is
successful at time t. Mathematically,
System up time
Availability =
System up time + System down time
MTTF
=
MTTF + MTTR
Availability is a measure of success used primarily for repairable systems. For
non-repairable systems, availability, A(t), equals reliability, R(t). In repairable
systems, A(t) will be equal to or greater than R(t).
The mean time between failures (MTBF) is an important measure in repairable
systems. This implies that the system has failed and has been repaired. Like MTTF
and MTTR, MTBF is an expected value of the random variable time between
failures. Mathematically,
MTBF = MTTF + MTTR
The term MTBF has been widely misused. In practice, MTTR is much smaller
than MTTF, which is approximately equal to MTBF. MTBF is often incorrectly
substituted for MTTF, which applies to both repairable systems and non-repairable
systems. If the MTTR can be reduced, availability will increase, and the system
will be more economical.
A system where faults are rapidly diagnosed is more desirable than a system
that has a lower failure rate but where the cause of a failure takes longer to detect,
resulting in a lengthy system downtime. When the system being tested is renewed
through maintenance and repairs, E(T) is also known as MTBF.
2.2 Common Distribution Functions

This section presents some of the common distribution functions and several
hazard models that have applications in reliability engineering (Pham 2000a).
Binomial Distribution
The binomial distribution is one of the most widely used discrete random variable
distributions in reliability and quality inspection. It has applications in reliability
engineering, e.g., when one is dealing with a situation in which an event is either a
success or a failure.
The pdf of the distribution is given by
()
P( X = x) = nx p x (1 − p) n − x x = 0, 1, 2, ..., n
( nx ) = x!(nn−! x)!
where n = number of trials; x = number of successes; p = single trial probability of
success.
The reliability function, R(k), (i.e., at least k out of n items are good) is given by
∑ ( nx ) p
n
R(k ) = x
(1 − p) n − x
x=k
Example 2.2: Suppose in the production of lightbulbs, 90% are good. In a random
sample of 20 lightbulbs, what is the probability of obtaining at least 18 good
lightbulbs?
Solution: The probability of obtaining 18 or more good lightbulbs in the sample of

20 is
20
⎛ 20 ⎞
R (18) = ∑
x =18
x
⎜ ⎟ (0.9) (0.1)
⎝ ⎠
18
20 − x
= 0.667
Poisson Distribution
Although the Poisson distribution can be used in a manner similar to the binomial
distribution, it is used to deal with events in which the sample size is unknown.
This is also a discrete random variable distribution whose pdf is given by
(λ t ) x e − λ t
P( X = x) = for x = 0, 1, 2, ....
x!
where λ = constant failure rate, x = is the number of events. In other words, P(X = x)
is the probability of exactly x failures occurring in time t. Therefore, the reliability
Poisson distribution, R(k) (the probability of k or fewer failures) is given by
(λ t ) x e − λ t
k
R(k ) = ∑
x=0 x!
This distribution can be used to determine the number of spares required for the
reliability of standby redundant systems during a given mission.
Exponential Distribution
Exponential distribution plays an essential role in reliability engineering because it
has a constant failure rate. This distribution has been used to model the lifetime of
electronic and electrical components and systems. This distribution is appropriate
when a used component that has not failed is as good as a new component - a
rather restrictive assumption. Therefore, it must be used diplomatically since
numerous applications exist where the restriction of the memoryless property may
not apply. For this distribution, we have reproduced equations (2.2) and (2.3),
respectively:
t
1 −
f (t ) = e θ = λ e−λt , t ≥ 0
θ t
−
R (t ) = e θ = e− λ t , t ≥ 0
where θ = 1/λ > 0 is an MTTF’s parameter and λ ≥ 0 is a constant failure rate.
The hazard function or failure rate for the exponential density function is
constant, i.e.,
1
1 −θ
e
f (t ) θ 1
h (t ) = = = =λ
R(t ) −
1 θ
e θ
The failure rate for this distribution is λ, a constant, which is the main reason
for this widely used distribution. Because of its constant failure rate property, the
exponential is an excellent model for the long flat “intrinsic failure” portion of the
bathtub curve. Since most parts and systems spend most of their lifetimes in this
portion of the bathtub curve, this justifies frequent use of the exponential (when
early failures or wear out is not a concern). The exponential model works well for
inter-arrival times. When these events trigger failures, the exponential lifetime
model can be used.
We will now discuss some properties of the exponential distribution that are
useful in understanding its characteristics, when and where it can be applied.
Property 2.1: (Memoryless property) The exponential distribution is the only

continuous distribution satisfying
P{T ≥ t} = P{T ≥ t + s | T ≥ s} for t > 0, s > 0 (2.8)
This result indicates that the conditional reliability function for the lifetime of
a component that has survived to time s is identical to that of a new component.
This term is the so-called "used-as-good-as-new" assumption.
The lifetime of a fuse in an electrical distribution system may be assumed to
have an exponential distribution. It will fail when there is a power surge causing
the fuse to burn out. Assuming that the fuse does not undergo any degradation over
time and that power surges that cause failure are likely to occur equally over time,
then use of the exponential lifetime distribution is appropriate, and a used fuse that
has not failed is as good as new.
Property 2.2: If T1, T2, ..., Tn, are independently and identically distributed
exponential random variables (RVs) with a constant failure rate λ, then
n
2λ ∑ Ti ~ χ 2 (2n)
i =1
where χ (2n) is a chi-squared distribution with degrees of freedom 2n. This result
2
is useful for establishing a confidence interval for λ.
Example 2.3: A manufacturer performs an operational life test on ceramic capaci-

tors and finds they exhibit constant failure rate with a value of 3x10-8 failure per
hour. What is the reliability of a capacitor at 104 hours?
Solution: The reliability of a capacitor at 104 hour is

−8 −4
R(t ) = e− λt = e−3×10 t
= e−3×10 = 0.9997
The resulting reliability plot is shown in Figure 2.1.
Normal Distribution
Normal distribution plays an important role in classical statistics owing to the
Central Limit Theorem. In reliability engineering, the normal distribution primarily
applies to measurements of product susceptibility and external stress. This two-
parameter distribution is used to describe systems in which a failure results due to
some wearout effect for many mechanical systems.
0.9
0.8
0.7
0.6
0 2000000 4000000 6000000 8000000 10000000
Figure 2.1. Reliability function vs time
The normal distribution takes the well-known bell shape. This distribution is
symmetrical about the mean and the spread is measured by variance. The larger the
value, the flatter the distribution. The pdf is given by
1 t −μ 2
1 − ( )
f (t ) = e 2 σ
-∞ < t < ∞
σ 2π
where μ is the mean value and σ is the standard deviation. The cumulative distri-
bution function (cdf) is
t 1 s−μ 2
1 − (
∫σ
)
F (t ) = e 2 σ ds
−∞ 2π
The reliability function is
∞ 1 s−μ 2
1 − (
R (t ) = ∫
)
e 2 σ ds
t σ 2π
There is no closed form solution for the above equation. However, tables for the
standard normal density function are readily available (see Table A1.1 in Appendix
1) and can be used to find probabilities for any normal distribution. If
T −μ
Z=
σ
is substituted into the normal pdf, we obtain
z2
1 −
f ( z) = e 2
-∞ < Z < ∞
2π
This is a so-called standard normal pdf, with a mean value of 0 and a standard
deviation of 1. The standardized cdf is given by
t 1
1 − s2
Φ (t ) =
−∞
∫ 2π
e 2 ds (2.9)
where Φ is a standard normal distribution function. Thus, for a normal random

variable T, with mean μ and standard deviation σ,
⎛ t−μ ⎞ ⎛t−μ ⎞
P(T ≤ t ) = P ⎜ Z ≤ = Φ⎜
⎝ σ ⎟⎠ ⎝ σ ⎠
⎟
where Φ yields the relationship necessary if standard normal tables are to be used.
The hazard function for a normal distribution is a monotonically increasing
function of t. This can be easily shown by proving that h’(t) ≥ 0 for all t. Since
f (t )
h (t ) =
R (t )
then (see Problem 15)
R (t ) f '(t ) + f 2 (t )
h '(t ) = ≥0 (2.10)
R 2 (t )
One can try this proof by employing the basic definition of a normal density
function f.
Example 2.4: A component has a normal distribution of failure times with μ =

2000 hours and σ = 100 hours. Find the reliability of the component and the hazard
function at 1900 hours.
Solution: The reliability function is related to the standard normal deviate z by

⎛ t−μ⎞
R (t ) = P ⎜ Z >
⎝ σ ⎟⎠
where the distribution function for Z is given by equation (2.9). For this particular
application,
⎛ 1900 − 2000 ⎞
R(1900) = P ⎜ Z > ⎟
⎝ 100 ⎠
= P ( z > −1)
From the standard normal table in Table A1.1 in Appendix 1, we obtain
R(1900) = 1 − Φ (−1) = 0.8413.
The value of the hazard function is found from the relationship
⎛ t−μ⎞
Φ⎜ z = ⎟
f (t ) σ ⎠
h (t ) = = ⎝
R (t ) σ R (t )
where φ is a pdf of standard normal density. Here
Φ(−1.0) 0.1587
h(1900) = =
σ R(t ) 100(0.8413)
= 0.0019 failures/cycle
Example 2.5: A part has a normal distribution of failure times with μ = 40000
cycles and σ = 2000 cycles. Find the reliability of the part at 38000 cycles.
Solution: The reliability at 38000 cycles

⎛ 38000 − 40000 ⎞
R(38000) = P ⎜ z > ⎟
⎝ 2000 ⎠
= P ( z > −1.0)
= Φ (1.0) = 0.8413
The resulting reliability plot is shown in Figure 2.2.
The normal distribution is flexible enough to make it a very useful empirical

model. It can be theoretically derived under assumptions matching many failure
mechanisms. Some of these are corrosion, migration, crack growth, and in general,
failures resulting from chemical reactions or processes. That does not mean that the
normal is always the correct model for these mechanisms, but it does perhaps
explain why it has been empirically successful in so many of these cases.
Log Normal Distribution

The log normal lifetime distribution is a very flexible model that can empirically fit
many types of failure data. This distribution, with its applications in
maintainability engineering, is able to model failure probabilities of repairable
systems and to model the uncertainty in failure rate information. The log normal
density function is given by
1 ⎛ ln t − μ ⎞
2
1 − ⎜ ⎟
2⎝ σ ⎠
f (t ) = e t≥0 (2.11)
σ t 2π
where μ and σ are parameters such that -∞ < μ < ∞, and σ > 0. Note that μ and σ
are not the mean and standard deviations of the distribution.
Reliability Curve
1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
35000 37000 39000 41000 43000 45000 47000 49000 51000 53000
Figure 2.2. Normal reliability plot vs time

The relationship to the normal (just take natural logarithms of all the data and
time points and you have “normal” data) makes it easy to work with many good
software analysis programs available to treat normal data.
Mathematically, if a random variable X is defined as X = lnT, then X is
normally distributed with a mean of μ and a variance of σ 2. That is,
E(X) = E(lnT) = μ
and
V(X) = V(lnT) = σ 2.
X
Since T = e , the mean of the log normal distribution can be found by using the
normal distribution. Consider that
⎡ 1 ⎛ x−μ ⎞ ⎤
2
∞ ⎢x − ⎜ ⎥
1 2 ⎝ σ ⎟⎠ ⎦⎥
E (T ) = E (e ) = ∫σ ⎣⎢
X
e dx
−∞ 2π
and by rearrangement of the exponent, this integral becomes
σ2 ∞ 1
μ+ 1 − [ x − ( μ +σ 2 )]2
E (T ) = e 2
∫σ
−∞ 2π
e 2σ 2
dx
Thus, the mean of the log normal distribution is

σ2
μ+
E (T ) = e 2
Proceeding in a similar manner,

E (T 2 ) = E (e 2 X ) = e2( μ +σ
2
)
thus, the variance for the log normal is

V (T ) = e 2 μ +σ (eσ − 1)
2 2
The cumulative distribution function for the log normal is

t 1 ln s − μ 2
1 − (
∫σs
)
F (t ) = e 2 σ ds
0 2π
and this can be related to the standard normal deviate Z by
F (t ) = P[T ≤ t ] = P(ln T ≤ ln t )
⎡ ln t − μ ⎤
= P ⎢Z ≤
⎣ σ ⎦⎥
Therefore, the reliability function is given by
⎡ ln t − μ ⎤
R (t ) = P ⎢ Z > (2.12)
⎣ σ ⎦⎥
and the hazard function would be
h (t ) =
f (t ) Φ σ
=
ln t − μ
( )
R(t ) σ t R (t )
where Φ is a cdf of standard normal density.
Example 2.6: The failure time of a certain component is log normal distributed
with μ = 5 and σ = 1. Find the reliability of the component and the hazard rate for a
life of 50 time units.
Solution: Substituting the numerical values of μ, σ, and t into equation (2.12), we

compute
⎡ ln 50 − 5 ⎤
R(50) = P ⎢ Z > ⎥ = P[ Z > −1.09]
⎣ 1 ⎦
= 0.8621
Similarly, the hazard function is given by
h(50) =
(
Φ ln 50 − 5
1 )
= 0.032 failures/unit.
50(1)(0.8621)
Thus, values for the log normal distribution are easily computed by using the
standard normal tables.
Example 2.7: The failure time of a part is log normal distributed with μ = 6 and σ =
2. Find the part reliability for a life of 200 time units.
Solution: The reliability for the part of 200 time units is

⎛ ln 200 − 6 ⎞
R(200) = P ⎜ Z > ⎟ = P( Z > −0.35)
⎝ 2 ⎠
= 0.6368
Reliability Curve
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 1000 2000 3000 4000 5000
Figure 2.3. Log normal reliability plot vs time
The log normal lifetime model, like the normal, is flexible enough to make it a
very useful empirical model. Figure 2.3 shows the reliability of the log normal vs
time. It can be theoretically derived under assumptions matching many failure
mechanisms. Some of these are: corrosion and crack growth, and in general,
failures resulting from chemical reactions or processes.
Weibull Distribution
The exponential distribution is often limited in applicability owing to the
memoryless property. The Weibull distribution (Weibull 1951) is a generalization
of the exponential distribution and is commonly used to represent fatigue life, ball
bearing life, and vacuum tube life. The Weibull distribution is extremely flexible
and appropriate for modeling component lifetimes with fluctuating hazard rate
functions and for representing various types of engineering applications. The
three-parameters probability density function is
β
⎛ t −γ ⎞
β (t − γ ) β −1 −⎝⎜ ⎟
f (t ) = e θ
⎠
t ≥γ ≥0
θβ
where θ and β are known as the scale and shape parameters, respectively, and γ
is known as the location parameter. These parameters are always positive. By using
different parameters, this distribution can follow the exponential distribution, the
normal distribution, etc. It is clear that, for t ≥ γ, the reliability function R(t) is
β
⎛ t −γ ⎞
−⎜ ⎟
R (t ) = e ⎝ θ ⎠
for t > γ > 0, β > 0, θ > 0 (2.13)
hence,
β (t − γ ) β −1
h (t ) = t > γ > 0, β > 0, θ > 0 (2.14)
θβ
It can be shown that the hazard function is decreasing for β < 1, increasing for
β > 1, and constant when β = 1.
Example 2.8: The failure time of a certain component has a Weibull distribution
with β = 4, θ = 2000, and γ = 1000. Find the reliability of the component and the
hazard rate for an operating time of 1500 hours.
Solution: A direct substitution into equation (2.13) yields

4
1500 −1000 ⎞
− ⎛⎜ ⎟
R(1500) = e = 0.996
⎝ 2000 ⎠
Using equation (2.14), the desired hazard function is given by

4(1500 − 1000)4 −1
h(1500) =
(2000) 4
= 3.13 x 10-5 failures/hour
Note that the Rayleigh and exponential distributions are special cases of the
Weibull distribution at β = 2, γ = 0, and β = 1, γ = 0, respectively. For example,
when β = 1 and γ = 0, the reliability of the Weibull distribution function in
equation (2.13) reduces to
t
−
R (t ) = e θ
and the hazard function given in equation (2.14) reduces to 1/θ, a constant. Thus,
the exponential is a special case of the Weibull distribution. Similarly, when γ = 0
and β = 2, the Weibull probability density function becomes the Rayleigh density
function. That is
t2
2 −
f (t ) = te θ for θ > 0, t ≥ 0
θ
Other Forms of Weibull Distributions

The Weibull distribution again is widely used in engineering applications. It was
originally proposed for representing the distribution of the breaking strength of
materials. The Weibull model is very flexible and also has theoretical justification
in many applications as a purely empirical model. Another form of Weibull
probability density function is, for example,
γ
f ( x) = λ γ xγ −1 e− λ t (2.15)
When γ=2, the density function becomes a Rayleigh distribution.
It can easily be shown that the mean, variance and reliability of the above
Weibull distribution are, respectively, as follows:
1
1
Mean = λ γ Γ(1 + )
γ
⎛ ⎛ 2 ⎞ ⎛ ⎛ 1 ⎞ ⎞2 ⎞
2
Variance = λ ⎜ Γ ⎜1 + ⎟ − ⎜ Γ ⎜1 + ⎟ ⎟ ⎟
γ
(2.16)
⎜ ⎝ γ ⎠ ⎝ ⎝ γ ⎠⎠ ⎟
⎝ ⎠
γ
Reliability = e−λt
1
Example 2.9: The time to failure of a part has a Weibull distribution with =250
λ
(measured in 105 cycles ) and γ=2. Find the part reliability at 106 cycles.
Solution: The part reliability at 106 cycles is
R(106 ) = e − (10) / 250 = 0.6703

2
The resulting reliability function is shown in Figure 2.4.

W e i b u l l R e l i a b i l i t y C u r ve
1
0 .9
0 .8
0 .7
0 .6
0 .5
0 .4
0 .3
0 .2
0 .1
0
0 10 20 30 40 50 60 70 80 90 100
Figure 2.4. Weibull reliability function vs time
Gamma Distribution
Gamma distribution can be used as a failure probability function for components
whose distribution is skewed. The failure density function for a gamma distribution
is
t
t α −1 −
f (t ) = α
e β t ≥ 0, α , β > 0 (2.17)
β Γ(α )
where α is the shape parameter and β is the scale parameter. Hence,
∞ s
1 −
R (t ) = ∫ s α −1
e β
ds
t
β α Γ (α )
If α is an integer, it can be shown by successive integration by parts that
− t α −1 ( βt )i
R (t ) = e β ∑
i =0 i!
(2.18)
and
− βt
f (t ) βα
1
Γ (α )
tα −1 e
h (t ) = =
R(t ) − βt
α −1 ( βt )i
e ∑
i =0 i!
The gamma density function has shapes that are very similar to the Weibull
distribution. At α = 1, the gamma distribution becomes the exponential distribu-
tion with the constant failure rate 1/β. The gamma distribution can also be used to
model the time to the nth failure of a system if the underlying failure distribution is
exponential. Thus, if Xi is exponentially distributed with parameter θ = 1/β, then
T = X1 + X2 +…+Xn, is gamma distributed with parameters β and n.
Example 2.10: The time to failure of a component has a gamma distribution with α
= 3 and β = 5. Determine the reliability of the component and the hazard rate at 10
time-units.
Solution: Using equation (2.18), we compute
(105 )
10 i
− 2
R(10) = e 5
∑ = 0.6767
i =0 i!
From equation (2.17), we obtain
f (10) 0.054
h(10) = = = 0.798 failures/unit time
R (10) 0.6767
The other form of the gamma probability density function can be written as
follows:
β α t α −1 −t β
f ( x) = e for t >0 (2.19)
Γ(α )
This pdf is characterized by two parameters: shape parameter α and scale
parameter β. When 0<α<1, the failure rate monotonically decreases; when α>1,
the failure rate monotonically increase; when α=1 the failure rate is constant.
The mean, variance and reliability of the density function in equation (2.19)
are, respectively,
α
Mean( MTTF) =
β
α
Variance = 2
β
∞ β α xα −1 − x β
Reliability = ∫ t Γ(α )
e dx
Example 2.11: A mechanical system time to failure is gamma distribution with

α=3 and 1/β=120. Find the system reliability at 280 hours.
Solution: The system reliability at 280 hours is given by

2
⎛ 280 ⎞
−280 2 ⎜ ⎟
⎝ 120 ⎠ = 0.85119
R(280) = e 120 ∑
k =0 k!
and the resulting reliability plot is shown in Figure 2.5.
The gamma model is a flexible lifetime model that may offer a good fit to
some sets of failure data. It is not, however, widely used as a lifetime distribution
model for common failure mechanisms. A common use of the gamma lifetime
model occurs in Bayesian reliability applications.
Reliability for Gamma Distribution

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Figure 2.5. Gamma reliability function vs time
Beta Distribution
The two-parameter Beta density function, f(t), is given by
Γ(α + β ) α
f (t ) = t (1 − t ) β 0 < t < 1, α > 0, β > 0
Γ(α )Γ( β )
where α and β are the distribution parameters. This two-parameter distribution is
commonly used in many reliability engineering applications.
Pareto Distribution
The Pareto distribution was originally developed to model income in a population.
Phenomena such as city population size, stock price fluctuations, and personal
incomes have distributions with very long right tails. The probability density
function of the Pareto distribution is given by
α kα
f (t ) = k ≤t≤∞
t α +1
The mean, variance and reliability of the Pareto distribution are, respectively,
Mean = k /(α − 1) for α > 1
αK2
Variance = for α > 2
(α − 1) 2 (α − 2)
α
⎛k⎞
Reliability = ⎜ ⎟
⎝t⎠
The Pareto and log normal distributions have been commonly used to model
the population size and economical incomes. The Pareto is used to fit the tail of the
distribution, and the log normal is used to fit the rest of the distribution.
Rayleigh Distribution
The Rayleigh function is a flexible lifetime distribution that can apply to many
degradation process failure modes. The Rayleigh probability density function is
⎛ −t 2 ⎞
⎜⎜ 2 ⎟⎟
t ⎝ 2σ ⎠
f (t ) = e (2.20)
σ2
The mean, variance, and reliability of Rayleigh function are, respectively,
1
⎛ π ⎞2
Mean = σ⎜ ⎟
⎝2⎠
⎛ π⎞
Variance = ⎜ 2 − ⎟σ 2
⎝ 2⎠
−σ t 2
Reliability = e 2
Example 2.12: Rolling resistance is a measure of the energy lost by a tire under
load when it resists the force opposing its direction of travel. In a typical car,
traveling at 60 miles per hour, about 20% of the engine power is used to overcome
the rolling resistance of the tires.
A tire manufacturer introduces a new material that, when added to the tire
rubber compound, significantly improves the tire rolling resistance but increases
the wear rate of the tire tread. Analysis of a laboratory test of 150 tires shows that
the failure rate of the new tire linearly increases with time (hours). It is expressed
as
h(t ) = 0.5 × 10−8 t
Find the reliability of the tire at one year.
Solution: The reliability of the tire after one year (8760 hours) of use is
0.5
− × 10−8 × (8760)2
R(1year ) = e 2 = 0.8254
Figure 2.6 shows the resulting reliability function.

Reliability Curv e
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 1 2 3 4 5 6 7 8 9 10
4
x 10
Figure 2.6. Rayleigh reliability function vs time
Vtub-shaped Hazard Rate Distribution

Pham (2002a) recently developed a two-parameter lifetime distribution with a
Vtub-shaped hazard rate, called Pham distribution - also known as Loglog
distribution.
Note that the loglog distribution with Vtub-shaped and Weibull distribution
with bathtub-shaped failure rates are not the same. As for the bathtub-shaped, after
the infant mortality period, the useful life of the system begins. During its useful
life, the system fails as a constant rate. This period is then followed by a wear out
period during which the system starts slowly and increases with the onset of wear
out. For the Vtub-shaped, after the infant mortality period, the system starts to
experience at a relatively low increasing rate, but this is not constant, and then
increases with failures due to aging.
The Pham probability density function is given as follows (Pham 2002a):
tα
f (t ) = α ln a t α −1 at e1− a
α
for t>0, a>0, α >0. (2.21)
The Pham distribution and reliability functions are
t tα
F (t ) = ∫ f ( x) dx = 1 − e1− a
0
and
tα
R(t ) = e1− a (2.22)
respectively. The corresponding failure rate of the Pham distribution is given by
α
h(t ) = α ln(a ) t α −1 a t (2.23)
Figures 2.7 and 2.8 describe the density function and failure rate function for
various values of a and α .
1
0.9
0.8
0.7 a=1.4 a=1.8
0.6 a=1.2
0.5 a=1
a=1.1
0.4
0.3 a=0.5
0.2 a=0.2
0.1
0
0 0.5 1 1.5 2 2.5 3 3.5 4
Figure 2.7. Probability density function for various values α with a=2
Figure 2.8. Probability density function for various values a with α = 1.5
Two-Parameter Hazard Rate Function

This is a two-parameter function that can have increasing and decreasing hazard
rates. The hazard rate, h(t), the reliability function, R(t), and the pdf are, respecti-
vely, given as follows
nλ t n −1
h (t ) = for n ≥ 1, λ > 0, t ≥ 0 (2.24)
λt n + 1
R(t ) = e− ln( λt
N
+1)
(2.25)
and
nλ t n −1 − ln( λt n
f (t ) = e +1)
n ≥ 1, λ > 0, t ≥ 0 (2.26)
λt n + 1
where n = shape parameter; λ = scale parameter
Three-Parameter Hazard Rate Function

This is a three-parameter distribution that can have increasing and decreasing
hazard rates. The hazard rate, h(t), is given as
λ (b + 1)[ln(λ t + α )]b
h (t ) = b ≥ 0, λ > 0, α ≥ 0, t ≥ 0 (2.27)
(λ t + α )
The reliability function R(t) for α = 1 is
R (t ) = e− ( ln (λ t + α ) )
b + 1
The probability density function f(t) is

b +1 λ (b + 1)[ln(λ t + α )]b
f (t ) = e−[ln(λ t +α )] (2.28)
(λ t + α )
where b = shape parameter, λ = scale parameter, and α = location parameter.
2.3 A Generalized Systemability Function

The traditional reliability definitions and its calculations have commonly been
carried out through the failure rate function within a controlled laboratory-test
environment. In other words, such reliability functions are applied to the failure
testing data and then utilized to make predictions on the reliability of the system
used in the field. The underlying assumption for such calculation is that the field
environments and the testing environments are the same.
By defintion, a mathematical reliability function is the probability that a system
will be successful in the interval from time 0 to time t, given by
t
∞
∫
− h ( s ) ds
R(t ) = ∫ f ( s )ds = e 0
(2.29)
t
where f(s) and h(s) are, respectively, the failure time density and failure rate
function.
The operating environments are, however, often unknown and yet different due
to the uncertainties of environments in the field (Pham and Xie 2003). A new look
at how reliability researchers can take account of the randomness of the field
environments into mathematical reliability modeling covering system failure in the
field is great interest.
Pham (2005a) recently developed a new mathematical function called system-
ability, considering the uncertainty of the operational environments in the function
for predicting the reliability of systems.
Notation
hi (t ) ith component hazard rate function
Ri (t ) ith component reliability function
λi Intensity parameter of Weibull distribution for ith component
λ λ = ( λ1 , λ2 , λ3 ..., λn ) .
γi Shape parameter of Weibull distribution for ith component
γ γ = ( γ 1 , γ 2 , γ 3 ..., γ n ) .
η A common environment factor
G (η ) Cumulative distribution function of η
α Shape parameter of Gamma distribution
β Scale parameter of Gamma distribution
2.3.1 Systemability Definition
This section discusses a definition of systemability function.
Definition 2.2 (Pham 2005a): Systemability is defined as the probability that the
system will perform its intended function for a specified mission time under the
random operational environments.
In a mathematical form, the systemabililty function is given by
t
∫
−η h ( s ) ds
Rs (t ) = ∫ e 0
dG (η ) (2.30)
η
where η is a random variable that represents the system operational environments

with a distribution function G.
This new function captures the uncertainty of complex operational environ-
ments of systems in terms of the system failure rate. It also would reflect the
reliability estimation of the system in the field.
If we assume that η has a gamma distribution with parameters α and β , i.e.,
η ~ gamma(α , β ) where the pdf of η is given by
β α xα −1e − β x
fη ( x) = for α , β >0; x ≥ 0 (2.31)
Γ(α )
then the systemability function of the system in equation (2.30), using the Laplace
transform (see Appendix 2), is given by
α
⎡ ⎤
⎢ ⎥
⎢ β ⎥
Rs (t ) = ⎢ t ⎥ (2.32)
⎢ β + h(s )ds ⎥
⎢⎣ 0
⎥⎦∫
2.3.2 Systemability Calculations
This subsection presents several systemability results and variances of some

system configurations such as series, parallel, and k-out-of-n systems (Pham
2005a). Consider the following assumptions:
1. A system consists of n independent components where the system is subject to a
random operational environment η.
2. ith component lifetime is assumed to follow the Weibull density function, i.e.
Component hazard rate hi (t ) = λi γ i t γ i −1 (2.33)
γi
Component reliability R i (t ) = e − λi t t>0 (2.34)
Given common environment factor η ~ gamma(α , β ) , the systemability functions

for different system structures can be obtained as follows.
Series System Configuration

In a series system, all components must operate successfully if the system is to
function. The conditional reliability function of series systems subject to an actual
operational random environment η is given by
⎛ n
⎞
⎜⎜ −η ∑ λi t γ i ⎟⎟
RSeries (t | η , λ , γ ) = e ⎝ i=1 ⎠
(2.35)
The series systemability is given as follows
α
⎡ ⎤
⎢ ⎥
⎛ n
⎞ β
∫ ⎝ i =1
∑
RSeries (t | λ , γ ) = exp ⎜ −η λi t γ i ⎟ dG (η ) = ⎢
⎠ ⎢ n
⎥
⎥
(2.36)
η
⎢
⎣
β +
i =1
λi t γ i ∑ ⎥
⎦
The variance of a general function R(t) is given by
Var [ R (t )] = E ⎡⎣ R 2 (t ) ⎤⎦ − ( E [ R(t ) ])
2
(2.37)
Given η ~ gamma(α , β ) , the variance of systemability for any system structure
can be easily obtained. Therefore, the variance of series systemability is given by
⎛ ⎛ n ⎞⎞
Var ⎡⎣ RSeries (t | λ , γ ) ⎤⎦ = ∫ exp ⎜⎜ −η ⎜ 2∑ λi t γ i ⎟ ⎟⎟dG (η ) −
η
⎝ ⎝ i =1 ⎠⎠
2
(2.38)
⎛ ⎛ n
⎞ ⎞
⎜⎜ ∫ η
⎝ i =1
∑
exp ⎜ −η λi t γ i ⎟ dG (η ) ⎟⎟
⎠
⎝ ⎠
or
α 2α
⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥
β β
Var ⎣⎡ RSeries (t | λ , γ ) ⎦⎤ = ⎢ ⎥ −⎢ ⎥ (2.39)
⎢ n
⎥ ⎢ n
⎥
⎣
∑
⎢ β + 2 λi t
i =1
γi
⎥
⎦
⎢ β + λi t
⎣ i =1
γi
∑ ⎥
⎦
Parallel System Configuration

A parallel system is a system that is not considered to have failed unless all
components have failed. The conditional reliability function of parallel systems
subject to the uncertainty operational environment η is given by
) ∑ exp ( −η ( λ t )) +
n
(
RParallel (t | η , λ , γ ) = exp −ηλi t γ i − i1
γ i1
+ λi t
2
γ i2
i1 ,i2 =1
i1 ≠ i2
( ( )) −
n
∑
γ i1 γ i2
exp −η λi t + λi t + λi t γ i 3
i1 ,i2 ,i2 =1
1 2 3
(2.40)
i1 ≠ i2 ≠ i3
.....
⎛ n
⎞
+ (−1) n −1 exp ⎜ −η ∑ λi t γ i ⎟
⎝ i =1 ⎠
Hence, the parallel systemability is given by
α α
n ⎡ ⎤
⎡ β n
⎤ β
R parallel (t | λ , γ ) = ⎢
i =1 ⎣ β
∑
+ λ t γi ⎥
− ⎢
⎢ β + λ t
γi γ
+ λi t i
∑ ⎥ +
⎥⎦
i ⎦ i ,i =1 ⎣ i 1 2
1 2
i1 ≠ i2 1 2
α
n ⎡ β ⎤
∑ ⎢
⎢ β + λ t
γi
+ λ
γi
+ λi t γ i
⎥ −
⎥⎦
i t
1 2 3
i ≠i ≠i ⎣
i ,i ,i =1
1 2 2 i
1 2 3
1 2 3
(2.41)
.....
α
⎡ ⎤
⎢ ⎥
β
+(−1) n −1 ⎢ ⎥
⎢ n
⎥
⎢ β + λi t
⎣ i =1
γi
∑ ⎥
⎦
or
α
⎡ ⎤
n ⎢ n
β ⎥
R parallel (t | λ , γ ) = (−1) ∑
k −1
⎢ ∑ γj ⎥
(2.42)
k =1 i1 , i2 ..., ik =1 ⎢ β +
i1 ≠ i2 ...≠ ik ⎢
∑ λ jt ⎥
⎣ j = i1 ,...ik ⎦⎥
To simplify the calculation of a general n-component parallel system, we only
consider here a parallel system consisting of two components. It is easy to see that
the second-order moments of the systemability function can be written as
γ1 γ
−2η λ 2 t 2 −2η ( λ 1 t γ 1 + λ 2 t γ 2 )
E ⎡⎣ R Parallel
2
(t | λ , γ ) ⎤⎦ = ∫ η (e−2η λ 1 t + e +e
γ γ γ γ γ
−η ( λ 1 t 1 + λ 2 t 2 ) −η (2 λ 1 t 1 + λ 2 t 2 ) −η ( λ 1 t γ 1 + 2 λ 2 t 2 )
+e −e − e ) d G (η )
The variance of series systemability of a two-component parallel system is

given by
α α
⎡ β ⎤ ⎡ β ⎤
Var ⎡⎣ RParallel (t | λ , γ ) ⎤⎦ = ⎢ γ1 ⎥
+⎢ γ2 ⎥
+
⎣ β + 2λ1t ⎦ ⎣ β + 2λ2 t ⎦
α α
⎡ β ⎤ ⎡ β ⎤
⎢ γ1 γ2 ⎥
+⎢ γ1 γ2 ⎥
−
⎣ β + 2λ1t + 2λ2 t ⎦ ⎣ β + λ1t + λ2 t ⎦
α α (2.43)
⎡ β ⎤ ⎡ β ⎤
⎢ γ1 γ2 ⎥
−⎢ γ1 γ2 ⎥
−
⎣ β + 2λ1t + λ2 t ⎦ ⎣ β + λ1t + 2λ2 t ⎦
α α α 2
⎡⎡ β ⎤ ⎡ β ⎤ ⎡ β ⎤ ⎤
⎢⎢ γ ⎥
+ ⎢ γ2 ⎥
− ⎢ γ1 γ2 ⎥
⎥
⎢⎣ ⎣ β + λ1t 1 ⎦ ⎣ β + λ2 t ⎦ ⎣ β + λ1t + λ2 t ⎦ ⎥⎦
k-out-of-n System Configuration

In a k-out-of-n configuration, the system will operate if at least k out of n
components are operating. To simplify the complexity of the systemability
function, we assume that all the components in the k-out-of-n systems are identical.
Therefore, for a given common environment η, the conditional reliability function
of a component is given by
γ
R(t | η , λ , γ ) = e −ηλt (2.44)
The conditional reliability function of k-out-of-n systems subject to the uncer-
tainty operational environment η can be obtained as follows:
n
⎛n⎞
Rk − out − of − n (t | η , λ , γ ) = ∑ ⎜ ⎟ e −η jλ t (1 − e −ηλ t )( n − j )
γ γ
(2.45)
j
j =k ⎝ ⎠
Note that
n− j
⎛ n − j ⎞ −ηλ tγ l
(1 − e−ηλ t )( n − j ) = ∑ ⎜
γ
⎟ (−e )
l =0 ⎝ l ⎠
The conditional reliability function of k-out-of-n systems, from equation (2.45),
can be rewritten as
n
⎛ n ⎞ n− j ⎛ n − j ⎞
Rk − out − of − n (t | η , λ , γ ) = ∑ ⎜ ⎟ ∑ ⎜ l −η ( j + l ) λ t γ
⎟(−1) e (2.46)
j = k ⎝ j ⎠ l =0 ⎝ l ⎠
Then if η ~ gamma(α , β ) then the k-out-of-n systemability is given by
α
n
⎛ n ⎞ n− j ⎛ n − j ⎞ ⎡ β ⎤
∑
R(T ,...,T ) ( t | λ , γ ) = ⎜ ⎟ ⎜ ∑ ⎟ ( −1)
l
⎢ ⎥ (2.47)
⎢⎣ β + λ ( j + l ) t
γ
j =k ⎝ j ⎠ l =0 ⎝ l ⎠ ⎥⎦
1 n
It can be easily shown that

n
⎛ n⎞ n
⎛n⎞
∑ ⎜⎝ i ⎟⎠ ∑ ⎜⎝ j ⎟⎠ e −η ( i + j ) λ t γ γ
Rk2− out − of − n (t | η , λ , γ ) = (1 − e −ηλ t )(2 n − i − j ) (2.48)
i=k j =k
Since
2n −i − j
⎛ 2n − i − j ⎞ −ηλtγ l
∑
γ
(1 − e−ηλ t )(2 n −i − j ) = ⎜ ⎟ (−e ) (2.49)
l =0 ⎝ l ⎠
we can rewrite equation (2.48), after several simplifications, as follows
n
⎛n⎞ n
⎛n⎞ 2n −i − j
⎛ 2n − i − j ⎞ −η (i + j + l )λ tγ
Rk2− out − of − n (t | η , λ , γ ) = ∑ ⎜⎝ i ⎟⎠∑ ⎜⎝ j ⎟⎠(−1) ∑
i=k j =k
l
l =0
⎜
⎝ l
⎟e
⎠
(2.50)
Therefore, the variance of k-out-of-n system systemability function is given by

2
Var ( Rk / n (t | λ , γ ) = ∫ Rk2/ n (t | η , λ , γ )dG (η ) − ⎡ ∫ Rk / n (t | η , λ , γ )dG (η ) ⎤
η ⎣⎢ η ⎦⎥
2
n
⎛n⎞ n
⎛ n⎞ 2n −i − j
⎛ 2n − i − j ⎞ l ⎛ β ⎞
=∑ ⎜ ⎟ ∑ ⎜ ⎟ ∑ ⎜ ⎟ (−1) ⎜ γ ⎟
(2.51)
i =k ⎝ i ⎠ j=k ⎝ j ⎠ l =0 ⎝ l ⎠ ⎝ β + (i + j + l )λ t ⎠
2
⎛ n ⎛ n ⎞ n− j ⎛ n − j ⎞ β ⎞ ⎞
2
l ⎛
- ⎜ ∑ ⎜ ⎟∑ ⎜ − ⎜ ⎟ ⎟
⎜ j = k ⎝ j ⎠ l = 0 ⎝ l ⎠⎟
( 1)
⎝ β + ( j + l )λ t γ ⎠ ⎟
⎝ ⎠
Example 2.13: Consider a k-out-of-n system where λ = 0.0001, γ = 1.5, n = 5,

and η ~ gamma(α , β ) . Calculate the systemability of various k-out-of-n system
configurations.
Solution: The systemability of generalized k-out-of-5 system configurations is

given as follows:
α
5
⎛ 5 ⎞ 5− j ⎛ 5 − j ⎞ ⎡ β ⎤
Rk − out − of − n (t | λ , γ ) = ∑ j
∑
⎜ ⎟ ⎜
j = k ⎝ ⎠ l =0 ⎝ l
⎟ (−1)
⎠
l
⎢ γ ⎥
⎣ β + λ ( j + l )t ⎦
(2.52)
Figures 2.9 and 2.10 show the reliability function (conventional reliability
function) and systemability function (equation 2.52) of a series system (here k=5)
for α = 2, β = 3 and for α = 2, β = 1 , respectively.
Figures 2.11 and 2.12 show the reliability and systemability functions of a
parallel system (here k=1) for α = 2, β = 3 and for α = 2, β = 1 , respectively.
Similarly, Figures 2.13 and 2.14 show the reliability and systemability functions of
a 3-out-of-5 system for α = 2, β = 3 and for α = 2, β = 1 , respectively.
0.8
Reliability
0.6
0.4
0.2
0
0 100 200 300 400 500
Time (hours)
Rseries(t|n) Rseries(t)
Figure 2.9. Comparisons of series system reliability vs systemability functions for α = 2 and
β =3
0.8
0.6
0.4
0.2
0
0 100 200 300 400 500
Rseries(t|n) Rseries(t)
Figure 2.10. Comparisons of series system reliability vs. systemability functions for α = 2
and β =1
0.96
Reliability
0.92
0.88
0.84
0.8
0 100 200 300 400 500
Time (hours)
Rpara(t|n) Rpara(t)
Figure 2.11. Comparisons of parallel system reliability vs systemability function for α = 2

and β =3
0.8
0.6
0.4
0.2
0
0 100 200 300 400 500
Rpara(t|n) Rpara(t)
Figure 2.12. Comparisons of parallel system reliability vs Systemability functions for α = 2

and β =1
0.8
Reliability
0.6
0.4
0.2
0
0 100 200 300 400 500
Time (hours)
Rkn(t|n) Rkn(t)
Figure 2.13. Comparisons of k-out-of-n system reliability vs. systemability functions for
α = 2 and β =3
0.8
Rkn(t)
0.6
Rkn(t|η)
0.4
0.2
0
0 100 200 300 400 500
Rkn(t|n) Rkn(t)
Figure 2.14. Comparisons of k-out-of-n system reliability vs. systemability functions for
α = 2 and β =1
Variance of Systemability Calculations

Assume λ = 0.00001, γ = 1.5, n = 3, k = 2, and η ~ gamma(α , β ) , Figures 2.15
and 2.16 shows the systemability and its confidence intervals of a 2-out-of-3
system (Pham 1993) for α = 2, β = 1 and α = 2, β = 2 , respectively.
Upper Bound
1
0.9
RK -out-of-n (t )
Reliability
0.8
Lower Bound
0.7
0.6
0 100 200 300 400 500
Time (hours)
Figure 2.15. A 2-out-of-3 systemability and its 95% confidence interval where α = 2, β = 1
Upper Bound
1
RK -out-of-n (t )
0.95
Reliability
Lower Bound
0.9
0.85
0.8
0 100 200 300 400 500
Time (hours)
Figure 2.16. A 2-out-of-3 systemability and its 95% confidence interval (α = 2, β = 2)
2.4 System Reliability with Multiple Failure Modes

This section discusses various reliability and optimization aspects of systems
subject to multiple types of failure. It is assumed that the system component states
are statistically independent and identically distributed. Networks of relays, diode
circuits, fluid flow valves, etc. are a few examples of systems having components
subject to failure in either open or closed modes.
The designations “closed mode” and “short mode” both appear in this section,
and we will use the two terms interchangeably. Redundancy can be used to en-
hance the reliability of a system without any change in the reliability of the
individual components that form the system. However, in a two-failure mode

problem, redundancy may either increase or decrease the system's reliability.
Therefore, adding components to the system may not increase the system
reliability.
The reliability of a system subject to two kinds of failure is calculated as
follows (Malon 1989):
System reliability = Pr{system works in both modes} (2.53)

= Pr{system works in open mode} - Pr{system fails in
closed mode}+ Pr{system fails in both modes)
When the open- and closed-mode failure structures are dual of one another, i.e.
Pr{system fails in both modes} = 0, then the system reliability given by equation
(2.53) becomes
System reliability = 1 - Pr{system fails in open mode}

- Pr{system fails in closed mode} (2.54)
Notation
q0 The open-mode failure probability of each component (p0 = 1 - q0)
qs The short-mode failure probability of each component (ps = 1 – qs)
⎣⎢ x ⎦⎥ The largest integer not exceeding x
* Implies an optimal value
2.4.1 Reliability Calculations
The Series System

Consider a series system consisting of n components. In this series system, any one
component failing in an open mode causes system failure in open mode whereas all
components of the system must malfunction in short mode for the system to fail in
closed mode.
The probabilities of system fails in open mode and fails in short mode are
F0 (n) = 1 − (1 − q0 )n
and
Fs (n) = qsn
respectively. From equation (2.54), the system reliability is
Rs (n) = (1 − q0 ) n − qsn (2.55)
where n is the number of identical and independent components. In a series
arrangement, reliability with respect to closed system failure increases with the
number of components, whereas reliability with respect to open system failure
decreases.
Theorem 2.1: Let q0 and qs be fixed. There exists an optimum number of

components, say n*, that maximizes the system reliability. If we define
⎛ q ⎞
log ⎜ 0 ⎟
n0 = ⎝ 1 − qs ⎠
⎛ q ⎞
log ⎜ s ⎟
⎝ 1 − q0 ⎠
then the system reliability, Rs(n*), is maximum for
⎧⎪ ⎢ n0 ⎥ + 1 if n0 is not an integer
n* = ⎨ ⎣ ⎦ (2.56)
⎪⎩n0 or n0 + 1 if n0 is an integer
Proof: The proof is left as an exercise for the reader (see Problem 2.17).
Example 2.14: A switch has two failure modes: fail-open and fail-short. The
probability of switch open-circuit failure and short-circuit failure are 0.1 and 0.2
respectively. A system consists of n switches wired in series. That is, given q0 =
0.1 and qs = 0.2. Then
⎛ 0.1 ⎞
log ⎜ ⎟
n0 = ⎝ 1 − 0.2 ⎠ = 1.4
⎛ 0.2 ⎞
log ⎜ ⎟
⎝ 1 − 0.1 ⎠
Thus, n* = ⎢⎣1.4 ⎥⎦ + 1 = 2. Therefore, when n* = 2 the system reliability Rs(n) =
0.77 is maximized.
The Parallel System

Consider a parallel system consisting of n components. For a parallel confi-
guration, all the components must fail in open mode or at least one component
must malfunction in short mode to cause the system to fail completely. The system
reliability is given by
R p (n) = (1 − qs )n − q0n (2.57)
where n is the number of components connected in parallel. In this case, (1 – qs)n
represents the probability that no components fail in short mode, and q0n represents
the probability that all components fail in open mode.
Theorem 2.2: If we define

⎛ q ⎞
log ⎜ s ⎟
n0 = ⎝ 1 − q0 ⎠ (2.58)
⎛ q ⎞
log ⎜ 0 ⎟
⎝ 1 − qs ⎠
then the system reliability Rp(n*) is maximum for

⎧⎪ ⎢ n0 ⎥ + 1 if n0 is not an integer
n* = ⎨ ⎣ ⎦ (2.59)
⎪⎩n0 or n0 + 1 if n0 is an integer.
Proof: The proof is left as an exercise for the reader (see Problem 2.18).
It is observed that, for any range of q0 and qs, the optimal number of parallel
components that maximizes the system reliability is one, if qs > q0 (see Problem
2.19). For most other practical values of q0 and qs the optimal number turns out to
be two. In general, the optimal value of parallel components can be easily obtained
using equation (2.58).
The Parallel-Series System

Consider a system of components arranged so that there are m subsystems
operating in parallel, each subsystem consisting of n identical components in
series. Such an arrangement is called a parallel-series arrangement. The compo-
nents are subject to two types of failure: failure in open mode and failure in short
mode.
The systems are characterized by the following properties:

1. The system consists of m subsystems, each subsystem containing n i.i.d. compo-
nents.
2. A component is either good, failed open, or failed short. Failed components
can never become good, and there are no transitions between the open and short
failure modes.
3. The system can be (a) good, (b) failed open (at least one component in each
subsystem fails open), or (c) failed short (all the components in any subsys-
tem fail short).
4.The unconditional probabilities of component failure in open and short modes are
known and are constrained: qo, qs > 0; qo + qs < 1.
The probabilities of a system failing in open mode and failing in short mode are
given by
F0 (m) = [1 − (1 − q0 ) n ]m (2.60)
and Fs (m) = 1 − (1 − (qs )n )m (2.61)

respectively. The system reliability is
R ps (n, m) = (1 − qsn ) m − [1 − (1 − q0 )n ]m (2.62)
An interesting example in Barlow and Proschan (1965) shows that there exists
no pair n, m maximizing system reliability, since Rps is made arbitrarily close to
one by appropriate choice of m and n. To see this, let
log qs − log(1 − q0 )
a= M n = qs− n /(1+ a ) mn = ⎣⎢ M n ⎦⎥
log qs + log(1 − q0 )
For given n, take m = mn; then one can rewrite equation (2.62) as:
R ps (n, mn ) = (1 − qsn )m − [1 − (1 − q0 ) n ]m
n n
A straightforward computation yields

lim R ps (n, mn ) = lim{(1 − qsn ) m − [1 − (1 − q0 ) n ]m } = 1
n n
n →∞ n →∞
For fixed n, q0, and qs, one can determine the value of m that maximizes Rps.
Theorem 2.3 (Barlow and Proschan 1965): Let n, q0, and qs be fixed. The
maximum value of Rps(m) is attained at m* = ⎢⎣ m0 ⎥⎦ + 1 , where
n(log p0 − log qs )
m0 = (2.63)
log(1 − qsn + log(1 − p0n )
If mo is an integer, then mo and mo+1 both maximize Rps(m).
Proof: The proof is left as an exercise for the reader (see Problem 20).
The Series-Parallel System

The series-parallel structure is the dual of the parallel-series structure. We consider
a system of components arranged so that there are m subsystems operating in
series, each subsystem consisting of n identical components in parallel. Such an
arrangement is called a series-parallel arrangement.
Failure in open mode of all the components in any subsystem makes the system
unresponsive. Failure in closed (short) mode of a single component in each
subsystem also makes the system unresponsive. The probabilities of system failure
in open and short mode are given by
F0 (m) = 1 − (1 − q0n ) m (2.64)
and
Fs (m) = [1 − (1 − qs ) n ]m (2.65)
respectively. The system reliability is
R(m) = (1 − q0n ) m − [1 − (1 − qs ) n ]m (2.66)
where m is the number of identical subsystems in series and n is the number of
identical components in each parallel subsystem.
Barlow and Proschan (1965) show that there exists no pair (m, n) maximizing
system reliability. For fixed m, q0, and qs however, one can determine the value of
n that maximizes the system reliability.
Theorem 2.4 (Barlow and Proschan 1965): Let n, q0, and qs be fixed. The
maximum value of R(m) is attained at m* = ⎢⎣ m0 ⎥⎦ + 1 , where
n(log ps − log q0 )
m0 = (2.67)
log(1 − q0n ) − log(1 − psn )
If mo is an integer, then mo and mo + 1 both maximize R(m).
Proof: (see Problem 21).
The k-out-of-n Systems

Consider a k-out-of-n system consisting of n identical and independent components
that can be either good or failed. The components are subject to two types of
failure: failure in open mode and failure in closed mode. The k out of n system can
fail when k or more components fail in closed mode or when (n - k + 1) or more
components fail in open mode.
Applications of k-out-of-n systems can be found in the areas of target detection,
communication, and safety monitoring systems, and, particularly, in the area of
human organizations. The following is an example in the area of human
organizations (Nordmann and Pham 1999).
Consider a committee with n members who must decide to accept or reject
innovation-oriented projects. The projects are of two types: "good" and "bad". It is
assumed that the communication among the members is limited, and each member
will make a yes-no decision on each project. A committee member can make two
types of error: the error of accepting a bad project and the error of rejecting a good
project. The committee will accept a project when k or more members accept it,
and will reject a project when (n - k + 1) or more members reject it.
Thus, the two types of potential error of the committee are: (1) the acceptance
of a bad project (which occurs when k or more members make the error of
accepting a bad project); (2) the rejection of a good project (which occurs when (n
- k + 1) or more members make the error of rejecting a good project).
This section determines the optimal k or n that maximizes the system reliability.
We also study the effect of the system's parameters on the optimal k or n. The
system fails in closed mode if and only if at least k of its n components fail in
closed mode, and we obtain
k −1
n
⎛n⎞ ⎛n⎞
Fs (k , n) = ∑ ⎜⎝ i ⎟⎠ q p
i =k
i
s
n −i
s = 1− ∑ ⎜⎝ i ⎟⎠q p
i =0
i
s
n −i
s (2.68)
The system fails in open mode if and only if at least (n - k +1) of its n components
fail in open mode, that is:
n
⎛ n ⎞ i n − i k −1 ⎛ n ⎞ i n −i
F0 (k , n) = ∑ ⎜ ⎟ q0 p0 = ∑ ⎜ ⎟ p0 q0
i = n − k +1 ⎝ i ⎠ i =0 ⎝ i ⎠
(2.69)
The system reliability is given by

k −1 k −1
⎛n⎞ ⎛n⎞
R(k , n) = 1 − F0 (k , n) − Fs (k , n) = ∑ ⎜⎝ i ⎟⎠ q p
i =0
i
s
n −i
s − ∑ ⎜⎝ i ⎟⎠ p q
i =0
i n −i
0 0 (2.70)
For a given k, we can find the optimum value of n, say n*, that maximizes the
system reliability.
Theorem 2.5 (Pham 1989a): For fixed k, q0, and qs, the maximum value of R(k,
n) is attained at n* = ⎢⎣ n0 ⎥⎦ where
⎡ ⎛ 1 − q0 ⎞ ⎤
⎢ log ⎜ ⎟⎥
n0 = k ⎢1 + ⎝ qs ⎠ ⎥ (2.71)
⎢ ⎛ 1 − qs ⎞ ⎥
⎢ log ⎜ ⎟⎥
⎢⎣ ⎝ q0 ⎠ ⎥⎦
If n0 is an integer, both n0 and n0 + 1 maximize R(k, n).
This result shows that when n0 is an integer, both n*-1 and n* maximize the system
reliability R(k, n). In such cases, the lower value will provide the more economical
optimal configuration for the system. If q0 = qs the system reliability R(k, n) is
maximized when n = 2k or 2k-1. In this case, the optimum value of n does not
depend on the value of q0 and qs and the best choice for a decision voter is a
majority voter; this system is also called a majority system (Pham,1989a).
From Theorem 2.5, we understand that the optimal system size n* depends on
the various parameters q0 and qs. It can be shown the optimal value n* is an
increasing function of q0 and a decreasing function of qs (see Problem 23).
Intuitively, these results state that when qs increases it is desirable to reduce the
number of components in the system as close to the value of threshold level k as
possible. On the other hand, when q0 increases, the system reliability will be
improved if the number of components increases.
Theorem 2.6 (Ben-Dov 1980): For fixed n, q0, and qs, it is straightforward to see
that the maximum value of R(k, n) is attained at k* = ⎣⎢ k0 ⎦⎥ +1, where
⎛q ⎞
log ⎜ 0 ⎟
k0 = n ⎝ ps ⎠ (2.72)
⎛qq ⎞
log ⎜ s 0 ⎟
⎝ ps p0 ⎠
If k0 is an integer, both k0 and k0 + 1 maximize R(k, n).
We now discuss how these two values, k* and n*, are related to one another.
Define α by
⎛q ⎞
log ⎜ 0 ⎟
α= ⎝ ps ⎠ (2.73)
⎛ qs q0 ⎞
log ⎜ ⎟
⎝ ps p0 ⎠
then, for a given n, the optimal threshold k is given by k* = ⎢⎡ nα ⎥⎤ and for a given k
the optimal n is n* = ⎣⎢ k / α ⎦⎥ . For any given q0 and qs, we can easily show that (see
Problem 25)
qs < α < p0 (2.74)

Therefore, we can obtain the following bounds for the optimal value of the
threshold k:
nqs < k * < np0 (2.75)
This result shows that for given values of q0 and qs, an upper bound for the optimal
threshold k* is the expected number of components working in open mode, and a
lower bound for the optimal threshold k* is the expected number of components
failing in closed mode.
2.4.2 An Application of Systems with Multiple Failure Modes
In many critical applications of digital systems, fault tolerance has been an

essential architectural attribute for achieving high reliability. Several techniques
can achieve fault tolerance using redundant hardware (Mathur and De Sousa 1975)
or software (Pham 1985).
Typical forms of redundant hardware structures for fault-tolerant systems are of
two types: fault masking and standby. Masking redundancy is achieved by
implementing the functions so that they are inherently error correcting, e.g.
triple-modular redundancy (TMR), N-modular redundancy (NMR), and self-
purging redundancy. In standby redundancy, spare units are switched into the
system when working units break down. Mathur and De Sousa (1975) have
analyzed, in detail, hardware redundancy in the design of fault-tolerant digital
systems. Redundant software structures for fault-tolerant systems based on the
acceptance tests have been proposed by Homing et al. (1974).
This section presents a fault-tolerant architecture to increase the reliability of a
special class of digital systems in communication (Pham and Upadhyaya 1989b).
In this system, a monitor and a switch are associated with each redundant unit. The
switches and monitors can fail. The monitors have two failure modes: failure to
accept a correct result, and failure to reject an incorrect result. The scheme can be
used in communication systems to improve their reliability.
Consider a digital circuit module designed to process the incoming messages
in a communication system. This module consists of two units: a converter to
process the messages, and a monitor to analyze the messages for their accuracy.
For example, the converter could be decoding or unpacking circuitry, whereas the
monitor could be checker circuitry (Lala 1985).
To guarantee a high reliability of operation at the receiver end, n converters
are arranged in "parallel". All, except converter n, have a monitor to determine if
the output of the converter is correct. If the output of a converter is not correct, the
output is cancelled and a switch is changed so that the original input message is
sent to the next converter. The architecture of such a system has been proposed by
Pham and Upadhyaya (1989b). Systems of this kind have useful applications in
communication and network control systems and in the analysis of fault-tolerant
software systems.
We assume that a switch is never connected to the next converter without a
signal from the monitor, and the probability that it is connected when a signal
arrives is ps. We next present a general expression for the reliability of the system
consisting of n non-identical converters arranged in "parallel". Let us define the

following notation, events, and assumptions.
Notation
pic Pr{converter i works}
pis Pr{switch i is connected to converter (i + 1) when a signal arrives}
m1
p i Pr{monitor i works when converter i works} = Pr{not sending a signal to
the switch when converter i works}
pim 2 Pr{i monitor works when converter i has failed} = Pr{sending a signal to
the switch when converter i has failed}
Rnk− k Reliability of the remaining system of size (n-k) given that the first k swit-
ches work
Rn Reliability of the system consisting of n converters
The events are:

Ciw , Ci f Converter i works, fails
M iw , M i f Monitor i works, fails
w f
S , Si
i Switch i works, fails
W System works
The assumptions are:

1. The system, the switches, and the converters are two-state: good or failed.
2. The module (converter, monitor, or switch) states are mutually statistically in-
dependent.
3. The monitors have three states: good, failed in mode 1, failed in mode 2.
4. The modules are not identical.
The reliability of the system is defined as the probability of obtaining the

correctly processed message at the output. To derive a general expression for the
reliability of the system, we use an adapted form of the total probability theorem as
translated into the language of reliability.
Let A denote the event that a system performs as desired. Let Xi and Xj be the
event that a component X (e.g. converter, monitor, or switch) is good or failed
respectively. Then
Pr{system works} = Pr{system works when unit X is good} x Pr{unit X is good}
+ Pr{system works when unit X fails} x Pr{unit X is failed}
The above equation provides a convenient way of calculating the reliability of

complex systems. Notice that R1 = pic and for n ≥ 2, the reliability of the system
can be calculated as follows:
Rn = Pr{W | C1w and M 1w } Pr{ C1w and M 1w } + Pr{W | C1w and M 1f }
Pr{ C1w and M 1f } + Pr{W | C1f and M 1w } Pr{ C1f and M 1w }
+ Pr{W | C1f and M 1f } Pr{ C1f and M 1f }

In order for the system to operate when the first converter works and the first
monitor fails, the first switch must work and the remaining system of size n-1 must
work:
Pr{W | C1w and M 1f } = p1s Rn1−1
Similarly,
Pr{W | C1f and M 1w } = p1s Rn1−1
then
Rn = p1c p1m1 + [ p1c (1 − p1m1 ) + (1 − p1c ) p1m2 ] p1s Rn1−1
The reliability of the system consisting of n non-identical converters can be
rewritten as:
n −1
Rn = ∑p
i =1
c
i pim1π i −1 + π n −1 pnc for n > 1 (2.76)
and R1 = p1c where

k
π kj = ∏ Ai for k ≥ 1
i= j
π k ≡ π k1 for all k and π 0 = 1

and
Ai ≡ [ pic (1 − pim1 ) + (1 − pic ) pim 2 ] for all i = 1, 2,...., n
Assume that all the converters, monitors, and switches have the same reliability,
that is
pic = p c , pim1 = p m1 , pim 2 = p m 2 , pis = p s for all i
then we obtain a closed form expression for the reliability of system as follows:
p c p m1
Rn = (1 − An −1 ) + p c An −1 (2.77)
1− A
where
A = [ p c (1 − p m1 ) + (1 − p c ) p m 2 ] p s
2.5 Markov Processes

Stochastic processes are used for the description of a systems operation over time.
There are two main types of stochastic processes: continuous and discrete. The
complex continuous process is a process describing a system transition from state
to state. The simplest process that will be discussed here is a Markov process.
Given the current state of the process, its future behavior does not depend on the
past. In Section 2.6 we will discuss the discrete stochastic process. As an intro-
duction to the Markov process, let us examine the following example.
Example 2.15: Consider a parallel system consisting of two components. From a

reliability point of view, the states of the system can be described by
State 1: Full operation (both components operating)
State 2: One component operating - one component failed
State 3: Both components failed
Define
Pi (t ) = P[ X (t ) = i ] = P[system is in state i at time t ]
and
Pi (t + dt ) = P[ X (t + dt ) = i ] = P[system is in state i at time t + dt ].
Define a random variable X(t) which can assume the values 1, 2, or 3
corresponding to the above-mentioned states. Since X(t) is a random variable, one
can discuss P[X(t) = 1], P[X(t) = 2] or conditional probability, P[X(t1)= 2 | X(t0)=
1]. Again, X(t) is defined as a function of time t, the last stated conditional
probability, P[X(t1)= 2 | X(t0)= 1], can be interpreted as the probability of being in
state 2 at time t1, given that the system was in state 1 at time t0. In this example, the
"stage space" is discrete, i.e., 1, 2, 3, etc., and the parameter space (time) is
continuous. The simple process described above is called a stochastic process, i.e.,
a process which develops in time (or space) in accordance with some probabilistic
(stochastic) laws. There are many types of stochastic processes. In this section, the
emphasis will be on Markov processes which are a special type of stochastic
process.
Definition 2.3: Let t0 < t1 < ... < tn. If

P[ X (tn ) = An | X (tn −1 ) = An −1 , X (tn − 2 ) = An − 2 ,...., X (t0 ) = A0 ]
= P[ X (tn ) = An | X (tn −1 ) = An −1 ]
then the process is called a Markov process.
Given the present state of the process, its future behavior does not depend on past
information of the process.
The essential characteristic of a Markov process is that it is a process that has no
memory; its future is determined by the present and not the past. If, in addition to
having no memory, the process is such that it depends only on the difference
(t+dt)-t = dt and not the value of t, i.e., P[X(t + dt) = j | X(t) = i] is independent of t,
then the process is Markov with stationary transition probabilities or homogeneous
in time. This is the same property noted in exponential event times, and referring
back to the graphical representation of X(t), the times between state changes would
in fact be exponential if the process has stationary transition probabilities.
Thus, a Markov process which is time homogeneous can be described as a
process where events have exponential occurrence times. The random variable of
the process is X(t), the state variable rather than the time to failure as in the
exponential failure density. To see the types of processes that can be described, a
review of the exponential distribution and its properties will be made. Recall that,
if X1, X2,.…, Xn, are independent random variables, each with exponential density
and a mean equal to 1/λi then min { X1, X2, … , Xn} has an exponential density with
(∑ λ )
−1
mean i .
The significance of the property is as follows:
1. The failure behavior of the simultaneous operation of components can be
characterized by an exponential density with a mean equal to the reciprocal of
the sum of the failure rates.
2. The joint failure/repair behavior of a system where components are operating
and/or undergoing repair can be characterized by an exponential density with a
mean equal to the reciprocal of the sum of the failure and repair rates.
3. The failure/repair behavior of a system such as 2 above, but further
complicated by active and dormant operating states and sensing and switching,
can be characterized by an exponential density.
The above property means that almost all reliability and availability models can be
characterized by a time homogeneous Markov process if the various failure times
and repair times are exponential. The notation for the Markov process is {X(t),
t>0}, where X(t) is discrete (state space) and t is continuous (parameter space). By
convention, this type of Markov process is called a continuous parameter Markov
chain.
From a reliability/availability viewpoint, there are two types of Markov processes.
These are defined as follows:
1. Absorbing Process: Contains what is called an "absorbing state" which is a
state from which the system can never leave once it has entered, e.g., a failure
which aborts a flight or a mission.
2. Ergodic Process Contains no absorbing states such that X(t) can move around
indefinitely, e.g., the operation of a ground power plant where failure only
temporarily disrupts the operation.
Pham (2000a) page 265, presents a summary of the processes to be considered
broken down by absorbing and ergodic categories. Both reliability and availability
can be described in terms of the probability of the process or system being in
defined "up" states, e.g., states 1 and 2 in the initial example. Likewise, the mean
time between failures (MTBF) can be described as the total time in the “up” states
before proceeding to the absorbing state or failure state.
Define the incremental transition probability as

Pij (dt ) = P[ X (t + dt ) = j | X (t ) = i ]
This is the probability that the process (random variable X(t)) will go to state i
during the increment t to (t+dt), given that it was in state i at time t. Since we are
dealing with time homogeneous Markov processes, i.e., exponential failure and
repair times, the incremental transition probabilities can be derived from an analy-
sis of the exponential hazard function. In Section 2.1, it was shown that the hazard
function for the exponential with mean 1/λ was just λ. This means that the limiting
(as dt → 0 ) conditional probability of an event occurrence between t and t + dt,
given that an event had not occurred at time t, is just λ, i.e.,
P[t < X < t + dt | X > t ]

h(t ) = lim =λ
dt → 0 dt
The equivalent statement for the random variable X(t) is
h(t )dt = P[ X (t + dt ) = j | X (t ) = i ] = λ dt
Now, h(t)dt is in fact the incremental transition probability, thus the Pij(dt) can be
stated in terms of the basic failure and/or repair rates.
Returning to Example 2.15, a state transition can be easily constructed showing
the incremental transition probabilities for process between all possible states (see
Figure.B.4, Pham 2000a)
State 1: Both components operating
State 2: One component up - one component down
State 3: Both components down (absorbing state)
The loops indicate the probability of remaining in the present state during the dt
increment
P11(dt) = 1 - 2λdt P12(dt) = 2λdt P13(dt) = 0
P21(dt) = 0 P22(dt) = 1 - λdt P23(dt) = λdt
P31(dt) = 0 P32(dt) = 0 P33(dt) = 1
The zeros on Pij, i > j, denote that the process cannot go backwards, i.e., this is not
a repair process. The zero on P13 denotes that in a process of this type, the
probability of more than one event (e.g., failure, repair, etc.) in the incremental
time period dt approaches zero as dt approaches zero.
Except for the initial conditions of the process, i.e., the state in which the
process starts, the process is completely specified by the incremental transition
probabilities. The reason for the latter is that the assumption of exponential event
(failure or repair) times allows the process to be characterized at any time t since it
depends only on what happens between t and (t+dt). The incremental transition
probabilities can be arranged into a matrix in a way which depicts all possible
statewide movements. Thus, for the parallel configurations,
⎡ 1 2 3 ⎤
⎢1 − 2λ dt 2λ dt 0 ⎥⎥
[ pij (dt )] = ⎢
⎢ 0 1 − λ dt λ dt ⎥
⎢ ⎥
⎣ 0 0 1 ⎦
for i, j = 1, 2, or 3. The matrix [Pij(dt)] is called the incremental, one-step transition
matrix. It is a stochastic matrix, i.e., the rows sum to 1.0. As mentioned earlier, this
matrix along with the initial conditions completely describes the process.
Now, [Pij(dt)] gives the probabilities for either remaining or moving to all the
various states during the interval t to t + dt, hence,
P1(t + dt) = (1 - 2λdt)P1(t)
P2(t + dt) = 2λdt P1(t)(1 - λdt)P2(t)
P3(t + dt) = λdtP2(t) + P3(t)
By algebraic manipulation, we have

[ P1 (t + dt ) − P1 (t )]
= −2λ P1 (t )
dt
[ P2 (t + dt ) − P2 (t )]
= 2λ P1 (t ) − λ P2 (t )
dt
[ P3 (t + dt ) − P3 (t )]
= λ P2 (t )
dt
Taking limits of both sides as dt → 0 , we obtain
P1’(t) = - 2λP1(t)
P2’(t) = 2λP1(t) - 2λP2(t)
P3’(t) = λP2(t)
The above system of linear first-order differential equations can be easily solved
for P1(t) and P2(t), and therefore, the reliability of the configuration can be
obtained:
2
R(t ) = ∑ Pi (t )
i =1
Actually, there is no need to solve all three equations, but only the first two as P3(t)
does not appear and also P3(t) = 1 – P1(t) – P2(t). The system of linear, first-order
differential equations can be solved by various means including both manual and
machine methods. For purposes here, the manual methods employing the Laplace
transform (see Appendix 2) will be used.
∞
L[ Pi (t )] = ∫ e − st Pi (t )dt = fi ( s )
0
∞
L[ Pi ' (t )] = ∫ e − st Pi ' (t )dt = s fi ( s ) − Pi (0)
0
The use of the Laplace transform will allow transformation of the system of linear,
first-order differential equations into a system of linear algebraic equations which
can easily be solved, and by means of the inverse transforms, solutions of Pi(t) can
be determined.
Returning to the example, the initial condition of the parallel configuration is
assumed to be “full-up” such that
P1(t = 0) = 1, P2(t = 0) = 0, P3(t = 0) = 0
transforming the equations for P’1(t) and P’2(t) gives
sf1 ( s ) − P1 (t ) |t = 0 = −2λ f1 ( s )
sf 2 ( s ) − P2 (t ) |t = 0 = 2λ f1 ( s ) − λ f 2 ( s )
Evaluating P1(t) and P2(t) at t = 0 gives
sf1 ( s ) − 1 = −2λ f1 ( s )
sf 2 ( s ) − 0 = 2λ f1 ( s ) − λ f 2 ( s )
from which we obtain
( s + 2λ ) f1 ( s ) = 1
−2λ f1 ( s ) + ( s + λ ) f 2 ( s ) = 0
Solving the above equations for f1(s) and f2(s), we have

1
f1 ( s ) =
( s + 2λ )
2λ
f 2 (s) =
[( s + 2λ )( s + λ )]
From Appendix 2 of the inverse Laplace transforms,
P1 (t ) = e −2 λt
P2 (t ) = 2e− λ t − 2e −2 λ t
R(t ) = P1 (t ) + P2 (t ) = 2e− λ t − e −2 λ t
The example given above is that of a simple absorbing process where we are
concerned about reliability If repair capability in the form of a repair rate μ were
added to the model, the methodology would remain the same with only the final
result changing. With a repair rate μ added to the parallel configuration, the
incremental transition matrix would be
⎡ 1 − 2λ dt 2λ dt 0 ⎤
[ Pij (dt )] = ⎢ μ dt 1 − (λ + μ )dt λ dt ⎥⎥
⎢
⎢⎣ 0 0 1 ⎥⎦
The differential equations would become
P1’(t) = - 2λP1(t) + μP2(t)
P2’(t) = 2λP1(t) – (λ +μ )P2(t)
and the transformed equations would become
( s + 2λ ) f1 ( s ) − μ f 2 ( s ) = 1
−2λ f1 ( s ) + ( s + λ + μ ) f 2 ( s ) = 0
Hence, we obtain
(s + λ + μ )
f1 ( s ) =
( s − s1 )( s − s2 )
2λ
f2 (s) =
( s − s1 )( s − s2 )
where
−(3λ + μ ) + (3λ + μ )2 − 8λ 2
s1 =
2
−(3λ + μ ) − (3λ + μ )2 − 8λ 2
s2 =
2
From Appendix 2, we obtain
( s1 + λ + μ )e− s1t ( s2 + λ + μ )e − s2t

P1 (t ) = +
( s1 − s2 ) ( s2 − s1 )
2λ e− s1t 2λ e− s2t
P2 (t ) = +
( s1 − s2 ) ( s2 − s1 )
Thus, the reliability of two-component in a parallel system is given by
R(t ) = P1 (t ) + P2 (t )
( s1 + 3λ + μ )e− s1t − ( s2 + 3λ + μ )e − s2t (2.78)
=
( s1 − s2 )
System Mean Time Between Failures

Another parameter of interest in absorbing Markov processes is the mean time
between failures (MTBF). Recalling the previous example of a parallel configura-
tion with repair, the differential equations P1'(t) and P2'(t) describing the process
were
P1’(t) = - 2λP1(t) + μP2(t)
P2’(t) = 2λP1(t) – (λ +μ )P2(t)
Integrating both sides of the above equations yields
∞ ∞ ∞
∫ P '1 (t )dt = −2λ ∫ P1 (t )dt + μ ∫ P2 (t )dt

0 0 0
∞ ∞ ∞
∫ P2 (t )dt = 2λ ∫ P1 (t )dt − (λ + μ )∫ P2 (t )dt

'
0 0 0
From Section 2.1,

∞
∫ R(t )dt = MTBF

0
Similarly,
∞
∫ P (t )dt =
0
1 mean time spent in state 1, and
∫ P (t )dt =
0
2 mean time spent in state 2
Designating these mean times as T1 and T2, respectively, we have

P1 (t )dt |∞0 = −2λT1 + μT2
P2 (t )dt |∞0 = 2λT1 − (λ + μ )T2
But P1(t) = 0 as t → ∞ and P1(t) = 1 for t = 0. Likewise, P2(t) = 0 as t → ∞ and
P2(t) = 0 for t = 0. Thus,
−1 = −2λT1 + μT2
0 = 2λT1 − (λ + μ )T2
or, equivalently,
⎡ −1⎤ ⎡ −2λ μ ⎤ ⎡T1 ⎤
⎢ ⎥=⎢ − (λ + μ ) ⎦⎥ ⎢⎣T2 ⎥⎦
⎣ 0 ⎦ ⎣ 2λ
Therefore,
(λ + μ ) 1
T1 = T2 =
2λ 2
λ
(λ + μ ) 1 (3λ + μ )
MTBF = T1 + T2 = + =
2λ 2 λ 2λ 2
The MTBF for non-maintenance processes is developed exactly the same way
as just shown. What remains under absorbing processes is the case for availability
for maintained systems. The difference between reliability and availability for
absorbing processes is somewhat subtle. A good example is that of a communica-
tion system where, if such a system failed temporarily, the mission would continue,
but, if it failed permanently, the mission would be aborted. Consider the following
cold-standby configuration consisting of two units: one main unit and one spare
unit (Pham 2000a):
State 1: Main unit operating - spare OK
State 2: Main unit out - restoration underway
State 3: Spare unit installed and operating
State 4: Permanent failure (no spare available)
The incremental transition matrix is given by (see Figure B.8 in Pham 2000a, for a
detailed state transition diagram)
⎡1 − λ dt λ dt 0 0 ⎤
⎢ 0 1 − μ dt μ dt 0 ⎥⎥
[ Pij (dt )] = ⎢
⎢ 0 0 1 − λ dt λ dt ⎥
⎢ ⎥
⎣ 0 0 0 1 ⎦
We obtain
P1’(t) = - λP1(t)
P2’(t) = λP1(t) - μP2(t)
P3’(t) = μP2(t) - λP3(t)
Using the Laplace transform, we obtain
sf1 ( s) − 1 = −λ f1 ( s)
sf 2 ( s) = λ f1 ( s) − μ f 2 ( s )
sf3 ( s) = μ f 2 ( s) − λ f 3 ( s )
After simplifications,
1
f1 ( s) =
(s + λ )
λ
f 2 ( s) =
[( s + λ )( s + μ )]
λμ
f3 ( s) =
[( s + λ )2 ( s + μ )]
Therefore, the probability of full-up performance, P1(t), is given by
P1 (t ) = e − λt
Similarly, the probability of the system being down and under repair, P2(t), is
⎡ λ ⎤ − μt
P2 (t ) = ⎢
λ − μ ⎥ e −e (
− λt
)
⎣ ( ) ⎦
and the probability of the system being full-up but no spare available, P3(t), is
⎡ λμ ⎤ − μ t
P3 (t ) = ⎢ 2 ⎥
[e − e − λt − (λ − μ )te− λ t ]
⎣ (λ − μ ) ⎦
Hence, the point availability, A(t), is given by
A(t ) = P1 (t ) + P3 (t )
If average or interval availability is required, this is achieved by
⎛1⎞ T ⎛1⎞ T
⎜ ⎟
⎝t ⎠
∫ 0
A(t )dt = ⎜ ⎟ ∫ [ P1 (t ) + P3 (t )]dt
⎝t⎠ 0
(2.79)
where T is the interval of concern.

With the above example, cases of the absorbing process (both maintained and
non-maintained) have been covered insofar as "manual" methods are concerned. In
general, the methodology for treatment of absorbing Markov processes can be
“packaged” in a fairly simplified form by utilizing matrix notation. Thus, for exam-
ple, if the incremental transition matrix is defined as follows:
⎡1- 2λ dt 2λ dt 0 ⎤
[ Pij (dt )] = ⎢⎢ μ dt 1 − (λ + μ )dt λ dt ⎥⎥
⎢⎣ 0 0 1 ⎥⎦
then if the dts are dropped and the last row and the last column are deleted, the
remainder is designated as the matrix T:
⎡1- 2λ 2λ ⎤
[T ] = ⎢
⎣ μ 1 − (λ + μ ) ⎥⎦
Define [Q] = [T]' - [I], where [T]' is the transposition of [T] and [I] is the unity
matrix:
⎡1 − 2λ μ ⎤ ⎡1 0⎤
[Q ] = ⎢ −
⎣ 2λ 1 − (λ + μ ) ⎥⎦ ⎢⎣0 1⎥⎦
⎡ −2 λ μ ⎤
=⎢
⎣ 2λ − (λ + μ ) ⎥⎦
Further define [P(t)] and [P’(t)] as column vectors such that
⎡ P (t ) ⎤ ⎡ P '(t ) ⎤
[ P1 (t )] = ⎢ 1 ⎥ , [ P '(t )] = ⎢ 1 ⎥
⎣ P2 (t ) ⎦ ⎣ P2 '(t ) ⎦
then
[ P '(t )] = [Q ][ P(t )]
At the above point, solution of the system of differential equations will produce
solutions to P1(t) and P2(t). If the MTBF is desired, integration of both sides of the
system produces
⎡ −1⎤ ⎡T1 ⎤
⎢ 0 ⎥ = [Q] ⎢T ⎥
⎣ ⎦ ⎣ 2⎦
⎡ −1⎤ ⎡ −2λ μ ⎤ ⎡T1 ⎤
⎢ 0 ⎥ = ⎢ 2λ − (λ + μ ) ⎥ ⎢T ⎥ or
⎣ ⎦ ⎣ ⎦⎣ 2⎦
⎡1 ⎤ ⎡T1 ⎤
[Q ]−1 ⎢ ⎥ = ⎢ ⎥
⎣ 0 ⎦ ⎣T2 ⎦
where [Q]-1 is the inverse of [Q] and the MTBF is given by
3λ + μ
MTBF = T1 + T2 =
(2λ ) 2
In the more general MTBF case,
⎡ −1⎤ ⎡ T1 ⎤
⎢0⎥ ⎢T ⎥
⎢ ⎥ ⎢ 2 ⎥
⎢ ⋅⎥ ⎢ ⋅ ⎥ n −1
[Q ]−1 ⎢ ⎥ = ⎢ ⎥ where ∑T i = MTBF
⎢ ⋅⎥ ⎢ ⋅ ⎥ i =1
⎢ ⋅⎥ ⎢ ⋅ ⎥
⎢ ⎥ ⎢ ⎥
⎢⎣ 0 ⎥⎦ ⎣⎢Tn −1 ⎦⎥
and (n - 1) is the number of non-absorbing states.
For the reliability/availability case, utilizing the Laplace transform, the system of
linear, first-order differential equations is transformed to
⎡ f ( s ) ⎤ ⎡1 ⎤ ⎡ f ( s) ⎤
s ⎢ 1 ⎥ − ⎢ ⎥ = [Q ] ⎢ 1 ⎥
⎣ f 2 ( s) ⎦ ⎣0 ⎦ ⎣ f 2 (s) ⎦
⎡ f1 ( s ) ⎤ ⎡1 ⎤
[ sI − Q] ⎢ ⎥=⎢ ⎥
⎣ f 2 ( s) ⎦ ⎣0 ⎦
⎡ f1 ( s ) ⎤ −1 ⎡1 ⎤
⎢ ⎥ = [ sI − Q] ⎢ ⎥
⎣ f 2 (s) ⎦ ⎣0 ⎦
⎡ f1 ( s ) ⎤ ⎡1 ⎤
L−1 ⎢ ⎥ = L−1{[ sI − Q ]−1 ⎢ ⎥}
⎣ f 2 ( s)⎦ ⎣ 0⎦
⎡ p1 ( s ) ⎤ −1 −1 ⎡1 ⎤
⎢ p ( s ) ⎥ = L {[ sI − Q] ⎢0 ⎥}
⎣ 2 ⎦ ⎣ ⎦
Generalization of the latter to the case of (n-1) non-absorbing states is
straightforward.
Ergodic processes, as opposed to absorbing processes, do not have any
absorbing states, and hence, movement between states can go on indefinitely For
the latter reason, availability (point, steady-state, or interval) is the only meaningful
measure. As an example for ergodic processes, a ground-based power unit config-
ured in parallel will be selected.
The parallel units are identical, each with exponential failure and repair times
with means 1/λ and 1/μ, respectively (Pham 2000a). Assume a two-repairmen
capability if required (both units down), then
State 1: Full-up (both units operating)
State 2: One unit down and under repair (other unit up)
State 3: Both units down and under repair
It should be noted that, as in the case of failure events, two or more repairs
cannot be made in the dt interval.
⎡ 1 − 2λ dt 2λ dt 0 ⎤
⎢
[ Pij (dt )] = ⎢ μ dt 1 − (λ + μ )dt λ dt ⎥⎥
⎣⎢ 0 2μ dt 1 − 2μ dt ⎦⎥
Case I: Point Availability - Ergodic Process. For an ergodic process, as t → ∞ the

availability settles down to a constant level. Point availability gives a measure of
things before the "settling down” and reflects the initial conditions on the process.
Solution of the point availability is similar to the case for absorbing processes
except that the last row and column of the transition matrix must be retained and
entered into the system of equations. For example, the system of differential
equations becomes
⎡ P1 '(t ) ⎤ ⎡ −2λ μ 0 ⎤ ⎡ P1 (t ) ⎤
⎢ ⎥ ⎢ ⎥ ⎢ P (t ) ⎥
⎢ P2 '(t ) ⎥ = ⎢ 2λ − (λ + μ ) 2μ ⎥⎢ 2 ⎥
⎢⎣ P3 '(t ) ⎥⎦ ⎢⎣ 0 λ − 2μ ⎥⎦ ⎢⎣ P3 (t ) ⎥⎦
Similar to the absorbing case, the method of the Laplace transform can be used to
solve for P1(t), P2(t), and P3(t), with the point availability, A(t), given by
A(t ) = P1 (t ) + P2 (t )
Case II: Interval Availability - Ergodic Process. This is the same as the absorbing
case with integration over time period T of interest. The interval availability, A(T), is
T
1
A(T ) =
T 0 ∫
A(t )dt (2.80)
Case III: Steady State Availability - Ergodic Process. Here the process is
examined as t → ∞ with complete “washout” of the initial conditions. Letting
t → ∞ the system of differential equations can be transformed to linear algebraic
equations. Thus,
⎡ P1 '(t ) ⎤ ⎡ −2λ μ 0 ⎤ ⎡ P1 (t ) ⎤
⎢ ⎥ ⎥ ⎢ P (t ) ⎥
lim ⎢ P2 '(t ) ⎥ = lim ⎢⎢ 2λ − (λ + μ ) 2μ ⎥⎢ 2 ⎥
t →∞ t →∞
⎢⎣ P3 '(t ) ⎥⎦ ⎢⎣ 0 λ − 2μ ⎥⎦ ⎢⎣ P3 (t ) ⎥⎦
As t → ∞ , Pi (t ) → constant and Pi '(t ) → 0 . This leads to an unsolvable sys-tem,

namely
⎡ 0 ⎤ ⎡ −2λ μ 0 ⎤ ⎡ P1 (t ) ⎤
⎢0 ⎥ = ⎢ 2λ − (λ + μ ) 2μ ⎥ ⎢ P (t ) ⎥
⎢ ⎥ ⎢ ⎥⎢ 2 ⎥
⎢⎣0 ⎥⎦ ⎢⎣ 0 λ − 2μ ⎥⎦ ⎢⎣ P3 (t ) ⎥⎦
To avoid the above difficulty, an additional equation is introduced:
3
∑ P (t ) = 1
i =1
i
With the introduction of the new equation, one of the original equations is deleted
and a new system is formed:
⎡1 ⎤ ⎡ 1 1 1 ⎤ ⎡ P1 (t ) ⎤
⎢ 0 ⎥ = ⎢ −2λ ⎢ ⎥
⎢ ⎥ ⎢ μ 0 ⎥⎥ ⎢ P2 (t ) ⎥
⎣⎢0 ⎦⎥ ⎣⎢ 2λ − (λ + μ ) 2μ ⎦⎥ ⎣⎢ P3 (t ) ⎦⎥
or, equivalently,
−1
⎡ P1 (t ) ⎤ ⎡ 1 1 1⎤ ⎡1 ⎤
⎢ P (t ) ⎥ = ⎢ −2λ μ 0 ⎥⎥ ⎢0 ⎥
⎢ 2 ⎥ ⎢ ⎢ ⎥
⎢⎣ P3 (t ) ⎥⎦ ⎢⎣ 2λ − (λ + μ ) 2μ ⎥⎦ ⎢⎣0 ⎥⎦
We now obtain the following results:

μ2
P1 (t ) =
( μ + λ )2
2λμ
P2 (t ) =
(μ + λ )2
and
P3 (t ) = 1 − P1 (t ) − P2 (t )
λ2
=
(μ + λ )2
Therefore, the steady state availability, A(∞), is given by
A3 (∞) = P1 (t ) + P2 (t )
μ ( μ + 2λ )
=
( μ + λ )2
Note that Markov methods can also be employed where failure or repair times
are not exponential, but can be represented as the sum of exponential times with
identical means (Erlang distribution or Gamma distribution with integer valued
shape parameters). Basically, the method involves the introduction of "dummy"
states which are of no particular interest in themselves, but serve the purpose of
changing the hazard function from constant to increasing.
2.6 Counting Processes

Among discrete stochastic processes, counting processes in reliability engineering
are widely used to describe the appearance of events in time, e.g., failures, number
of perfect repairs, etc. The simplest counting process is a Poisson process. The
Poisson process plays a special role to many applications in reliability (Pham
2000a). A classic example of such an application is the decay of uranium.
Radioactive particles from nuclear material strike a certain target in accordance
with a Poisson process of some fixed intensity. A well-known counting process is
the so-called renewal process. This process is described as a sequence of events,
the intervals between which are independent and identically distributed random
variables. In reliability theory, this type of mathematical model is used to describe
the number of occurrences of an event in the time interval. In this section we also
discuss the quasi-renewal process and the non-homogeneous Poisson process.
A non-negative, integer-valued stochastic process, N(t), is called a counting
process if N(t) represents the total number of occurrences of the event in the time
interval [0, t] and satisfies these two properties:
1. If t1 < t2, then N(t1) ≤ N(t2)
2. If t1 < t2, then N(t2) - N(t1) is the number of occurrences of the event in the
interval [t1 , t2]
For example, if N(t) equals the number of persons who have entered a restaurant at
or prior to time t, then N(t) is a counting process in which an event occurs
whenever a person enters the restaurant.
2.6.1 Poisson Processes
One of the most important counting processes is the Poisson process.
Definition 2.4: A counting process, N(t), is said to be a Poisson process with

intensity λ if
1. The failure process, N(t), has stationary independent increments
2. The number of failures in any time interval of length s has a Poisson
distribution with mean λs, that is,
e − λ s (λ s ) n
P{N (t + s ) − N (t ) = n} = n = 0,1, 2,... (2.81)
n!
3. The initial condition is N(0) = 0
This model is also called a homogeneous Poisson process indicating that the failure
rate λ does not depend on time t. In other words, the number of failures occurring
during the time interval (t, t + s] does not depend on the current time t but only the
length of time interval s. A counting process is said to possess independent incre-
ments if the number of events in disjoint time intervals are independent.
For a stochastic process with independent increments, the auto-covariance
function is
⎧Var[ N (t1 + s ) − N (t2 )] for 0 < t2 − t1 < s
Cov[ X (t1 ), X (t2 )] = ⎨
⎩0 otherwise
where
X (t ) = N (t + s ) − N (t ).
If X(t) is Poisson distributed, then the variance of the Poisson distribution is
⎧λ[ s − (t2 − t1 )] for 0 < t2 − t1 < s
Cov[ X (t1 ), X (t2 )] = ⎨
⎩0 otherwise
This result shows that the Poisson increment process is covariance stationary. We
now present several properties of the Poisson process.
Property 2.3: The sum of independent Poisson processes, N1(t), N2(t), …., Nk(t),
with mean values λ1t, λ2t, …., λkt respectively, is also a Poisson process with mean
⎛ k ⎞
∑
⎜ λi ⎟ t . In other words, the sum of the independent Poisson processes is also a
⎝ i =1 ⎠
Poisson process with a mean that is equal to the sum of the individual Poisson
process' mean.
Property 2.4: The difference of two independent Poisson processes, N1(t), and
N2(t), with mean λ1t and λ2t, respectively, is not a Poisson process. Instead, it has
the probability mass function
k
− ( λ1 + λ2 ) t ⎛ λ1 ⎞ 2
P[ N1 (t ) − N 2 (t ) = k ] = e ⎜ ⎟ I k (2 λ1λ2 t ) (2.82)
⎝ λ2 ⎠
where Ik(.) is a modified Bessel function of order k (Handbook 1980).
Property 2.5: If the Poisson process, N(t), with mean λt, is filtered such that every
occurrence of the event is not completely counted, then the process has a constant
probability p of being counted. The result of this process is a Poisson process with
mean λpt [ ].
Property 2.6: Let N(t) be a Poisson process and Yn a family of independent and
identically distributed random variables which are also independent of N(t). A
stochastic process X(t) is said to be a compound Poisson process if it can be
represented as
N (t )
X (t ) = ∑Y
i =1
i
2.6.2 Renewal Processes
A renewal process is a more general case of the Poisson process in which the
inter-arrival times of the process or the time between failures do not necessarily
follow the exponential distribution. For convenience, we will call the occurrence of
an event a renewal, the inter-arrival time the renewal period, and the waiting time
the renewal time.
Definition 2.5: A counting process N(t) that represents the total number of
occurrences of an event in the time interval (0, t] is called a renewal process, if the
time between failures are independent and identically distributed random variables.
The probability that there are exactly n failures occurring by time t can be written
as
P{N (t ) = n} = P{N (t ) ≥ n} − P{N (t ) > n} (2.83)
Note that the times between the failures are T1, T2, …, Tn so the failures occurring
at time Wk are
k
Wk = ∑T
i =1
i
and
Tk = Wk − Wk −1
Thus,
P{N (t ) = n} = P{N (t ) ≥ n} − P{N (t ) > n}
= P{Wn ≤ t} − P{Wn +1 ≤ t}
= Fn (t ) − Fn +1 (t )
where Fn(t) is the cumulative distribution function for the time of the nth failure
and n = 0,1,2, ... .
Example 2.16: Consider a software testing model for which the time to find an
error during the testing phase has an exponential distribution with a failure rate of
X. It can be shown that the time of the nth failure follows the gamma distribution
with parameters k and n with probability density function. From equation (2.83) we
obtain
P{N (t ) = n} = P{N (t ) ≤ n} − P{N (t ) ≤ n − 1}

n
(λ t ) k − λ t n −1 (λ t )k − λ t
= ∑
k =0 k!
e − ∑
k =0 k!
e
(λ t ) n − λ t
= e for n = 0,1, 2,....
n!
Several important properties of the renewal function are given below.
Property 2.7: The mean value function of the renewal process, denoted by m(t), is
equal to the sum of the distribution function of all renewal times, that is,
m(t ) = E[ N (t )]
∞
= ∑ F (t )
n =1
n
Proof: The renewal function can be obtained as

m(t ) = E[ N (t )]
∞
= ∑ nP{N (t ) = n}
n =1
∞
= ∑ n[ F (t ) −F
n =1
n n +1 (t )]
∞
= ∑ F (t )
n =1
n
The mean value function of the renewal process is also called the renewal function.
Property 2.8: The renewal function, m(t), satisfies the following equation:
∫
m(t ) = Fa (t ) + m(t − s )dFa ( s )
0
(2.84)
where Fa(t) is the distribution function of the inter-arrival time or the renewal
period. The proof is left as an exercise for the reader (see Problem 28).
In general, let y(t) be an unknown function to be evaluated and x(t) be any

non-negative and integrable function associated with the renewal process. Assume
that Fa(t) is the distribution function of the renewal period. We can then obtain the
following result.
Property 2.9: Let the renewal equation be

t
∫
y (t ) = x(t ) + y (t − s )dFa ( s )
0
(2.85)
then its solution is given by

t
∫
y (t ) = x(t ) + x(t − s )dm( s )
0
where m(t) is the mean value function of the renewal process.

The proof of the above property can be easily derived using the Laplace
transform. It is also noted that the integral equation given in Property 2.8 is a
special case of Property 2.9.
Example 2.17: Let x(t) = a. Thus, in Property 2.9, the solution y(t) is given by
t
∫
y (t ) = x(t ) + x(t − s )dm( s )
0
t
∫
= a + a dm( s )
0
= a(1 + E[ N (t )])
2.6.3 Quasi-renewal Processes
In this section, a general renewal process, namely, the quasi-renewal process, is

discussed. Let {N(t), t > 0} be a counting process and let Xn be the time between
the (n-1)th and the nth event of this process, n ≥ 1.
Definition 2.6 (Wang and Pham 1996a): If the sequence of non-negative random
variables {X1, X2, ....} is independent and
X i = aX i −1 (2.86)
for i ≥ 2 where α > 0 is a constant, then the counting process {N(t), t ≥ 0} is said to
be a quasi-renewal process with parameter and the first inter-arrival time X1.
When α = 1, this process becomes the ordinary renewal process as discussed

in Section 2.6.2. This quasi-renewal process can be used to model reliability
growth processes in software testing phases and hardware burn-in stages for α > 1,
and in hardware maintenance processes when α ≤ 1.
Assume that the probability density function, cumulative distribution function,
survival function, and failure rate of random variable X1 are f1(x), F1(x), s1(x), and
r1(x), respectively. Then the pfd, cdf, survival function, failure rate of Xn for n = 1,
2, 3,... is respectively given below (Wang and Pham 1996a):
1 ⎛ 1 ⎞
f n ( x) = f1 ⎜ n −1 x ⎟
α n −1⎝α ⎠
⎛ 1 ⎞
Fn ( x) = F1 ⎜ n −1 x ⎟
⎝α ⎠
⎛ 1 ⎞
sn ( x) = s1 ⎜ n −1 x ⎟
⎝α ⎠
1 ⎛ 1 ⎞
f n ( x) = n −1 r1 ⎜ n −1 x ⎟
α ⎝α ⎠
Similarly, the mean and variance of Xn is given as
E ( X n ) = α n −1 E ( X 1 )
Var ( X n ) = α 2 n − 2Var ( X 1 )
Because of the non-negativity of X1 and the fact that X1 is not identically 0, we
obtain
E ( X 1 ) = μ1 ≠ 0
Proposition 2.1 (Wang and Pham 1996a): The shape parameters of Xn are the
same for n = 1, 2, 3, ... for a quasi-renewal process if X1 follows the gamma,
Weibull, or log normal distribution.
This means that after "renewal", the shape parameters of the inter-arrival time
will not change. In software reliability, the assumption that the software debugging
process does not change the error-free distribution type seems reasonable. Thus,
the error-free times of software during the debugging phase modeled by a
quasi-renewal process will have the same shape parameters. In this sense, a
quasi-renewal process is suitable to model the software reliability growth. It is
worthwhile to note that
E ( X 1 + X 2 + ... + X n ) μ (1 − α n )
lim = lim 1
n →∞ n n →∞ (1 − α ) n
= 0 if α < 1
= ∞ if α > 1
Therefore, if the inter-arrival time represents the error-free time of a software

system, then the average error-free time approaches infinity when its debugging
process is occurring for a long debugging time.
Distribution of N(t)
Consider a quasi-renewal process with parameter α and the first inter-arrival time
X1. Clearly, the total number of renewals, N(t), that has occurred up to time t and
the arrival time of the nth renewal, SSn, has the following relationship:
N(t) ≥ n if and only if SSn ≤ t
that is, N(t) is at least n if and only if the nth renewal occurs prior to time t. It is
easily seen that
n n
SSn = ∑ X i =∑ α i −1 X 1 for n ≥ 1 (2.87)
i =1 i =1
Here, SS0 = 0. Thus, we have

P{N (t ) = n} = P{N (t ) = n} − P{N (t ) ≥ n + 1}
= P{SSn ≤ t} − P{SSn +1 ≤ t}
= Gn (t ) − Gn +1 (t )
where Gn(t) is the convolution of the inter-arrival times F1, F2, F3, …, Fn. In other
words,
Gn (t ) = P{F1 + F2 + .... + Fn ≤ t}
If the mean value of N(t) is defined as the renewal function m(t), then,
m(t ) = E[ N (t )]
∞
= ∑ P{N (t ) ≥ n}
n =1
∞
= ∑ P{SSn ≤ t}
n =1
∞
= ∑ Gn (t )
n =1
The derivative of m(t) is known as the renewal density

λ (t ) = m '(t )
In renewal theory, random variables representing the inter-arrival distributions
only assume non-negative values, and the Laplace transform of its distribution
F1(t) is defined by
∞
L {F1 ( s )} = ∫ e − sx dF1 ( x)
0
Therefore,
∞
LFn ( s ) = ∫ e − a
n −1
st
dF1 (t ) = LF1 (α n −1 s )
0
and
∞
Lmn ( s ) = ∑ LGn ( s )
n =1
∞
= ∑ LF1 ( s )LF1 (α s ) ⋅⋅⋅⋅ LF1 (α n −1 s )
n =1
Since there is a one-to-one correspondence between distribution functions and its

Laplace transform, it follows that
Proposition 2.2 (Wang and Pham 1996a): The first inter-arrival distribution of a
quasi-renewal process uniquely determines its renewal function.
If the inter-arrival time represents the error-free time (time to first failure), a
quasi-renewal process can be used to model reliability growth for both software
and hardware.
Suppose that all faults of software have the same chance of being detected. If
the inter-arrival time of a quasi-renewal process represents the error-free time of a
software system, then the expected number of software faults in the time interval
[0, t] can be defined by the renewal function, m(t), with parameter α > 1. Denoted
by mr(t), the number of remaining software faults at time t, it follows that
mr (t ) = m(Tc ) − m(t )
where m(Tc) is the number of faults that will eventually be detected through a
software lifecycle Tc.
2.6.4 Non-homogeneous Poisson Processes
The non-homogeneous Poisson process model (NHPP) that represents the number
of failures experienced up to time t is a non-homogeneous Poisson process {N(t), t
≥ 0}. The main issue in the NHPP model is to determine an appropriate mean value
function to denote the expected number of failures experienced up to a certain
time.
With different assumptions, the model will end up with different functional
forms of the mean value function. Note that in a renewal process, the exponential
assumption for the inter-arrival time between failures is relaxed, and in the NHPP,
the stationary assumption is relaxed.
The NHPP model is based on the following assumptions:
• The failure process has an independent increment, i.e., the number of failures
during the time interval (t, t + s) depends on the current time t and the length
of time interval s, and does not depend on the past history of the process.
• The failure rate of the process is given by
P{exactly one failure in (t , t + Δt )} = P{N (t + Δt ) − N (t ) = 1}
= λ (t )Δt + o(Δt )
where λ(t) is the intensity function.
• During a small interval Δt, the probability of more than one failure is
negligible, that is,
P{two or more failure in (t , t + Δt )} = o(Δt )

• The initial condition is N(0) = 0.
On the basis of these assumptions, the probability of exactly n failures occurring

during the time interval (0, t) for the NHPP is given by
[m(t )]n − m (t )
Pr{N (t ) = n} = e n = 0,1, 2,... (2.88)
n!
t
where m(t ) = E[ N (t )] = ∫ λ ( s )ds and λ(t) is the intensity function. It can be easily
0
shown that the mean value function m(t) is non-decreasing.
Reliability Function
The reliability R(t), defined as the probability that there are no failures in the time
interval (0, t), is given by
R(t ) = P{N (t ) = 0}
= e− m(t )
In general, the reliability R(x|t), the probability that there are no failures in the
interval (t, t + x), is given by
R( x | t ) = P{N (t + x) − N (t ) = 0}
= e −[ m (t + x ) − m (t )]
and its density is given by
f ( x) = λ (t + x)e−[ m (t + x ) − m (t )]
where
∂
λ ( x) = [m( x)]
∂x
The variance of the NHPP can be obtained as follows:
t
Var[ N (t )] = ∫ λ ( s)ds
0
and the auto-correlation function is given by

Cor[ s ] = E[ N (t )]E[ N (t + s ) − N (t )] + E[ N 2 (t )]
t t+s t
= ∫ λ ( s )ds ∫ λ ( s )ds + ∫ λ ( s )ds

0 0 0
t
⎡ t+s
⎤
= ∫ λ ( s )ds ⎢1 + ∫ λ (s)ds ⎥⎥⎦
0 ⎢⎣ 0
Example 2.18: Assume that the intensity λ is a random variable with the pdf f(λ).
Then the probability of exactly n failures occurring during the time interval (0, t) is
given by
∞
(λ t ) n
P{N (t ) = n} = ∫ e− λ t f (λ ) d λ
0
n!
It can be shown that if the pdf f(λ) is given as the following gamma density
function with parameters k and m,
1
f (λ ) = k m λ m −1e − k λ for λ ≥ 0
Γ ( m)
then
⎛ n + m − 1⎞ m n
P{N (t ) = n} = ⎜ ⎟ p q n = 0,1, 2,...
⎝ n ⎠
is also called a negative binomial density function, where
k t
p= and q = = 1− p
t+k t+k
2.7 Further Reading

The reader interested in a deeper understanding of advanced probability theory and
stochastic processes should note the following highly recommended books:
Devore, J.L., Probability and Statistics for Engineering and the Sciences, 3rd
edition, Brooks/Cole Pub. Co., Pacific Grove, 1991.
Gnedenko, BV and I.A. Ushakov, Probabilistic Reliability Engineering, Wiley,
New York, 1995.
Feller, W., An Introduction to Probability Theory and Its Applications, 3rd edition,
Wiley, New York, 1994.
2.8 Problems
1. Assume that the hazard rate, h(t), has a positive derivative. Show that the
hazard distribution
t
H (t ) = ∫ h( x)dx
0
is strictly convex.
2. An operating unit is supported by (n-1) identical units on cold standby. When

it fails, a unit from standby takes its place. The system fails if all n units fail.
Assume that units on standby cannot fail and the lifetime of each unit follows
the exponential distribution with failure rate λ .
(a) What is the distribution of the system lifetime?
(b) Determine the reliability of the standby system for a mission of 100
hours when λ = 0.0001 per hour and n = 5.
3. Assume that there is some latent deterioration process occurring in the

system. During the interval [0, a-h] the deterioration is comparatively small
so that the shocks do not cause system failure. During a relatively short time
interval [a-h, a], the deterioration progresses rapidly and makes the system
susceptible to shocks. Assume that the appearance of each shock follows the
exponential distribution with failure rate λ . What is the distribution of the
system lifetime?
4. Consider a series system of n Weibull components. The corresponding

lifetimes T1, T2,..., Tn are assumed to be independent with pdf
⎪⎧λ β β t β −1e i
β
−(λ t )
for t ≥ 0
f (t ) = ⎨ i
⎪⎩0 otherwise
where λ > 0 and β > 0 are the scale and shape parameters, respectively.
(a) Show that the lifetime of a series system has the Weibull distribution
with pdf
⎧ n ⎛ n
⎞β
⎞ β −1 −⎜⎜⎝ ∑ λ
β
⎪⎛ λ β ⎟⎟ t
∑
i
⎟ βt e for t ≥ 0
⎠
f s (t ) = ⎨⎜⎝ i =1 i
i =1
⎠
⎪
⎩0 otherwise
(b) Find the reliability of this series system.
5. Consider the pdf of a random variable that is equally likely to take on any
value only in the interval from a to b.
(a) Show that this pdf is given by
⎧ 1
⎪ for a < t < 0
f (t ) = ⎨ b − a
⎪⎩0 otherwise
(b) Derive the corresponding reliability function R(t) and failure rate h(t).
(c) Think of an example where such a distribution function would be of
interest in reliability application.
6. The failure rate function, denoted by h(t), is defined as

d
h (t ) = − ln[ R(t )]
dt
Show that the constant failure rate function implies an exponential

distribution.
7. One thousand new streetlights are installed in Saigon city. Assume that the
lifetimes of these streetlights follow the normal distribution. The average life
of these lamps is estimated at 980 burning-hours with a standard deviation of
100 hours.
(a) What is the expected number of lights that will fail during the first 800
burning-hours?
(b) What is the expected number of lights that will fail between 900 and
1100 burning-hours?
(c) After how many burning-hours would 10% of the lamps be expected to
fail?
8. A fax machine with constant failure rate λ will survive for a period of 720
hours without failure, with probability 0.80.
(a) Determine the failure rate λ .
(b) Determine the probability that the machine, which is functioning after
600 hours, will still function after 800 hours.
(c) Find the probability that the machine will fail within 900 hours, given
that the machine was functioning at 720 hours.
9. The time to failure T of a unit is assumed to have a log normal distribution

with pdf
(ln t − μ ) 2
1 1 −
f (t ) = e 2σ 2
t>0
2π σ t
Show that the failure rate function is unimodal.
10. A diode may fail due to either open or short failure modes. Assume that the
time to failure T0 caused by open mode is exponentially distributed with pdf
f 0 (t ) = λ0 e − λ t
0
t≥0
and the time to failure T1 caused by short mode has the pdf
f s (t ) = λs e− λ t
s
t≥0
The pdf for the time to failure T of the diode is given by
f (t ) = p f 0 (t ) + (1 − p) f s (t ) t≥0
(a) Explain the meaning of p in the above pdf function.
(b) Derive the reliability function R(t) and failure rate function h(t) for the
time to failure T of the diode.
(c) Show that the diode with pdf f(t) has a decreasing failure rate (DFR).
11. A diesel is known to have an operating life (in hours) that fits the following
pdf:
2a
f (t ) = t≥0
(t + b ) 2
The average operating life of the diesel has been estimated to be 8760 hours.
(a) Determine a and b.
(b) Determine the probability that the diesel will not fail during the first
6000 operating-hours.
(c) If the manufacturer wants no more than 10% of the diesels returned for
warranty service, how long should the warranty be?
12. The failure rate for a hydraulic component

t
h (t ) = t >0
t +1
where t is in years.
(a) Determine the reliability function R(t).
(b) Determine the MTTF of the component.
13. A 18-month guarantee is given based on the assumption that no more than 5%
of new cars will be returned.
(a) The time to failure T of a car has a constant failure rate. What is the
maximum failure rate that can be tolerated?
(b) Determine the probability that a new car will fail within three years
assuming that the car was functioning at 18 months.
14. Show that if

R1 (t ) ≥ R2 (t ) for all t
where Ri(t) is the system reliability of the structure i, then MTTF of the system
structure 1 is always ≥ MTTF of the system structure 2.
15. Prove equation (2.10)
16. Show that the reliability function of Pham distribution (see equation 2.21) is
given as in equation (2.22).
17. Prove Theorem 2.1.
19. Show that for any range of q0 and qs, if qs > q0, the optimal number of parallel
components that maximizes the system reliability is one.

23. Show that the optimal value n* in Theorem 2.5 is an increasing function of q0
and a decreasing function of qs.
25. For any given q0 and qs, show that qs < α < p0 where α is given in equation
(2.73).
26. Prove Property 2.3.
29. Events occur according to an NHPP in which the mean value function is
m(t) = t3 + 3t2 + 6t t > 0.
What is the probability that n events occur between times t = 10 and t = 15?
http://www.springer.com/978-1-85233-950-0

c1

Uploaded by

Copyright:

Available Formats

c1

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

c1

Uploaded by

Copyright:

Available Formats

2

System Reliability Concepts

The analysis of the reliability of a system must be based on precisely defined

2.1 Reliability Measures

Solution: From equation (2.1) we obtain

System Mean Time to Failure

Thus, MTTF is the definite integral evaluation of the reliability function. In

Table 2.1. Results of a twelve-component life duration test

Component Time to failure

Failure Rate Function

For example, if g (t ) = μ e− μ t where μ > 0 is a constant repair rate, then

2.2 Common Distribution Functions

Solution: The probability of obtaining 18 or more good lightbulbs in the sample of

Property 2.1: (Memoryless property) The exponential distribution is the only

is useful for establishing a confidence interval for λ.

Example 2.3: A manufacturer performs an operational life test on ceramic capaci-

Solution: The reliability of a capacitor at 104 hour is

Figure 2.1. Reliability function vs time

where Φ is a standard normal distribution function. Thus, for a normal random

Example 2.4: A component has a normal distribution of failure times with μ =

Solution: The reliability function is related to the standard normal deviate z by

Solution: The reliability at 38000 cycles

The normal distribution is flexible enough to make it a very useful empirical

Log Normal Distribution

Figure 2.2. Normal reliability plot vs time

Thus, the mean of the log normal distribution is

Proceeding in a similar manner,

thus, the variance for the log normal is

The cumulative distribution function for the log normal is

Solution: Substituting the numerical values of μ, σ, and t into equation (2.12), we

Solution: The reliability for the part of 200 time units is

Figure 2.3. Log normal reliability plot vs time

Solution: A direct substitution into equation (2.13) yields

Using equation (2.14), the desired hazard function is given by

Other Forms of Weibull Distributions

Solution: The part reliability at 106 cycles is

R(106 ) = e − (10) / 250 = 0.6703

The resulting reliability function is shown in Figure 2.4.

Figure 2.4. Weibull reliability function vs time

Solution: Using equation (2.18), we compute

Example 2.11: A mechanical system time to failure is gamma distribution with

Solution: The system reliability at 280 hours is given by

Reliability for Gamma Distribution

Figure 2.5. Gamma reliability function vs time

Figure 2.6 shows the resulting reliability function.

Vtub-shaped Hazard Rate Distribution

Two-Parameter Hazard Rate Function

Three-Parameter Hazard Rate Function

The probability density function f(t) is

2.3 A Generalized Systemability Function

2.3.1 Systemability Definition

This section discusses a definition of systemability function.

where η is a random variable that represents the system operational environments

2.3.2 Systemability Calculations

This subsection presents several systemability results and variances of some

Given common environment factor η ~ gamma(α , β ) , the systemability functions

Series System Configuration

Parallel System Configuration