See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/302505024

Handbook of Performability Engineering

Book · July 2008

DOI: 10.1007/978-1-84800-131-2

CITATIONS READS

316 2,244

1 author:

Krishna B. Misra
RAMS Consultants
132 PUBLICATIONS 3,466 CITATIONS

SEE PROFILE

All content following this page was uploaded by Krishna B. Misra on 19 June 2017.

The user has requested enhancement of the downloaded file.

20

Tampered Failure Rate Load-Sharing Systems: Status and

Perspectives

Suprasad V. Amari1, Krishna B. Misra2, and Hoang Pham3

1
Relex Software Corporation, U.S.A.
2
RAMS Consultants, Jaipur, India.
3
Department of Industrial Engineering, Rutgers University, U.S.A.

Abstract: Load-sharing systems have several practical applications. In load-sharing systems, the event of
a component failure will result in a higher load, therefore inducing a higher failure rate, in each of the
surviving components. This introduces failure dependency among the load-sharing components, which in
turn increases the complexity in analyzing these systems. In this chapter, we first discuss the modelling
approaches and existing solution methods for analyzing the reliability of load-sharing systems. We then
describe tampered failure rate (TFR) load-sharing systems and their properties. Using these properties, we
provide efficient solution methods for solving TFR load-sharing models. Because load-sharing k-out-of-n
systems have several practical applications in reliability engineering, we provide a detailed analysis for
various cases of these systems. The solution methods proposed in this chapter are applicable for both
identical and non-identical component cases. The proposed methods are not restricted to the exponential
failure distributions and are applicable for a wide range of failure time distributions of the components. In
most cases, we provide closed-form analytical solutions for the reliability of TFR load-sharing k-out-of-
n:G systems. As a special case, efficient solutions are provided for systems with identical components
where all surviving components share the load equally. The efficiencies of the proposed methods are
demonstrated through several numerical examples.

20.1 Introduction many systems are load-sharing, where the

assumption of independence is no longer valid. In a
In reliability engineering, it is a common practice load-sharing system, if a component fails, the same
to use redundancy techniques to improve system workload has to be shared by the remaining
reliability [1]. In most cases, when analyzing components, resulting in an increased load shared
redundancy, independence is assumed across the by each surviving component. In most
components within the system. In other words, it is circumstances, an increased load induces a higher
assumed that the failure of a component does not component failure rate [2]. Many empirical studies
affect the failure properties (failure rates) of the of mechanical systems [3] and computer systems
remaining components. In the real-world, however, [4-5] have proved that the workload strongly
290 Suprasad Amari, K. B. Misra, and H. Pham

affects the component failure rate. Applications of sharing models have interesting applications in
load-sharing systems include electric generators nuclear reactor safety, software reliability [14],
sharing an electrical load in a power plant, CPUs in distributed computing [15], population sampling
a multiprocessor computer system, cables in a [8], combat modeling [16], modeling the
suspension bridge, and valves or pumps in a incubation period for the Human
hydraulic system [6]. Therefore, it is important to Immunodeficiency Virus (HIV) [17], and
develop reliability models that incorporate condition-based maintenance [18]. For a summary
stochastic dependencies among the system’s of these applications, refer to Kvam & Peña [8].
components. Even though load-sharing systems have been
The stochastic dependency models can be studied for a long time and have a wide range of
broadly classified as shock models and load- applications, the methods that are applicable for
sharing models [7-8]. In the shock models, the studying the time-dependent reliability
system is exposed to shocks that cause random characteristics of these systems are limited.
amounts of damage [9]. The shocks themselves can Majority of research papers related to this topic are
arrive according to a random process. The intensity published in material science, physics, and applied
and occurrence frequency of the shocks may vary statistics journals. These papers focus on the
with time. Generally, the occurrences of shocks are statistical properties of materials (strength of
modeled using homogeneous or non-homogeneous materials) subjected to load-sharing rules such as
Poisson processes. The additional damage to the equal load-sharing, local load-sharing, monotone
system at a given shock may depend on the load-sharing. The research publications that
intensity of the shock, damage that already consider both time-dependent failures (also failure
experienced, and the age of the system. The system rates) and the dynamic effects of loads are limited
fails when the cumulative damage exceeds a [2]. In this chapter, we present existing methods
certain level. An example of a shock model is the and their useful extensions for analyzing time-
failure of a dam due to excessive water in the dependent reliability characteristics of load-sharing
reservoir after several successive rainfalls [10]. systems. We focus our attention on a specific class
Another class of shock models includes common- of load-sharing models called tampered failure rate
cause failures. For example, the bivariate shock (TFR) load-sharing models [19]. Whenever
model introduced by Marshall-Olkin [11] analyzes applicable, we also present simplified results for
component dependencies by incorporating latent some special cases that include exponential failure
variables to allow simultaneous component distributions and identically distributed
failures. In the load-sharing models, the component components.
failure rates depend on the operating status of the Section 20.2 discusses the background and basic
other system components and the effective system concepts in modeling load-sharing system. Section
structure function. In 1945, Daniels [12] originally 20.3 presents a brief overview of static and
adopted the load-sharing model to describe how the dynamic methods. It also discusses the related
strain on yarn fibers increases as individual fibers works that are used for analyzing the time-
within a bundle break. For example, a bundle of dependent reliability characteristics of load-sharing
fibers can be considered as a parallel system systems. Section 20.4 presents the system
subject to a steady tensile load. In this chapter, we description, assumptions, and details of the
concentrate on the load-sharing models. tampered failure rate (TFR) models. Section 20.5
Load-sharing models have a wide range of presents the reliability analysis of load-sharing k-
engineering applications. They are extensively out-of-n systems with exponential failure
used in the textile engineering (fibers) [12], distributions, and section 20.6 focuses on the
material science and testing (fatigue and crack general failure distributions. Section 20.7 presents
growth) [13], mechanical engineering, and civil & the conclusions and future directions of research.
structural engineering (welded joint on large
support structures) [8] disciplines. Further, load-
 Tampered Failure Rate Load-Sharing Systems: Status and Perspectives 291

20.2 Basics of Load-Sharing Systems the stresses uniformly after individual

failures.
In order to analyze the reliability of load-sharing • Local load-sharing rule: A load on a failed
systems, we should consider the pattern of loads component is transferred to adjacent
acting on the system, load-sharing polices that components, and the proportion of the load
dictate how the load on the system is distributed that the surviving components inherit
among the components, and the relationship depends on their ‘distance’ to the failed
between the dynamic load and the failure behavior component. Examples of this kind include
of the components over a time period. cables supporting bridges and other
structures, composite materials with
20.2.1 Load Pattern bounding matrix joins, and transmission
systems that are modelled using consecutive
In load-sharing systems, the components of the k-out-of-n systems.
system are subjected to varying loads due to • Monotone load-sharing rule: The load on any
numerous reasons including changing demands of individual component is non-decreasing as
the system, failure of some components, changes in other components fail. This is a
the operating conditions, etc. Therefore, the total generalization to the previous two load-
load on the system can be categorized as: sharing rules. This load-sharing rule is
• Constant total load: In this case, the total load applicable if we are not redistributing any
on the system is constant. However, the portion of a load from a good component.
components experience different loads due to The same rules are also applicable for time-varying
the unavailability of some components as a loading conditions. It should be noted that for the
result of failure or preventive maintenance. 1-out-of-2 systems subjected to constant total load,
The constant total load assumption is all of these load-sharing rules produce the same
applicable for a wide range of applications. effects (or coincide).
The majority of published works are based on A generalization to the above load-sharing rules
this assumption. In this chapter, we also focus includes the non-monotone load-sharing rule,
on this assumption. where the load on working components may
• Time-varying total load: The total load on the increase or decrease depending on the dynamics in
system may vary with time. The variation can the systems. A simple example of this kind
be deterministic or random. Therefore, the includes load on a processor due to the variations
components can experience varying loads due in the demand [15]. The non-monotone load-
to the variations in the demand. Hence, the sharing rule may also be applicable (even if the
load on a component not only changes when total load on the system is constant) when the
other components fail but also when system components have discrete load-carrying capacities
demands fluctuate. (power generators). Similarly, a modification to the
equal load-sharing rule to analyze twisted fibre
20.2.2 Load-Sharing Rule bundles is proposed in [21].

The most important element of the load-sharing 20.2.3 Load-Life Relationship

model is the rule that governs how the loads on the
working components change after some In order to analyze the reliability of load-sharing
components in the system fail [20]. systems, we should consider the relationship
between the load and the failure behavior of a
• Equal load-sharing rule: A constant system component. In general, failure rate of a component
load is distributed equally among the working
increases with the applied load. The accelerated
components. Examples of this kind include
life testing models play an important role in
yarn bundles and untwisted cables that spread
determining the relationships between load and
292 Suprasad Amari, K. B. Misra, and H. Pham

failure rate. There is a much literature available on • Power Law: φ ( L) = Lα

the accelerated life testing models [22].
• Exponential Law: φ ( L) = e Lα
20.2.3.1 Proportional Hazards Model (PHM)
If the baseline distribution is Weibull (or
The PHM was first introduced by Cox [23] and has Exponential) and the multiplicative factor
recently gained popularity in the engineering field (acceleration factor) follows the power law, then
[24-26]. The PHM assumes that the hazard the AFTM and PHM coincide. However, in
(failure) rate of a component is the product of both general, there is no direct duality between the
a baseline hazard rate (which can be a function of models in all cases [28].
time t) and a multiplicative factor based on the If the failure distribution is Exponential and the
values of a set of conditions (loads in this case). In multiplicative factor follows the power law, then:
the PHM, component failure rate is expressed as: λ ( L) = λ ( L0 )( L / L0 )α (20.6)
h(t ; z ) = h0 (t ) ⋅ exp( zβ ) (20.1)
where z = {z 1 , …, z m } is a set of conditions (loads) 20.2.4 Effects of Load History on Life
acting on the component, and h 0 (t) is the baseline
failure rate function for the standard set of In the previous section, we discussed how the load
conditions (at the baseline load). In fact, exp(zβ) effects the failure distribution of a component.
can be replaced by any known function g(z,β). These models are applicable when a constant, fixed
When there is only one type of load, z = L, it load is acting on a component. However, in load-
reduces to: sharing systems, a component operates at different
h(t ; L) = h0 (t ) ⋅ exp( Lβ ) (20.2) loads during different time intervals. Therefore, it
If the failure rate of the component at a fixed load is important to consider the effects of load history
is a constant function of time, then we have: h 0 (t) on the component failure rate. Hence, we should
= λ 0 = λ(L 0 ). Then, we have: also consider other details that are applicable for
λ (t; L) = λ ( L) = λ0 ⋅ exp( Lβ ) (20.3) step-stress acceleration life testing models, where
the component experiences different loads at
When the baseline failure distribution is different time intervals. However, unlike in
Weibull or Exponential, under certain situations, accelerated life testing, in a load-sharing system,
the PHM model is also equivalent to the the load varies at random time intervals. Hence,
Accelerated Failure Time Model (AFTM) [27]. analyzing load-sharing systems is much more
20.2.3.2 Accelerated Failure Time Model (AFTM) complex than analyzing fixed duration step-stress
acceleration life testing models.
The AFTM is first proposed by Pike [27] and has In this section, we present the step-stress
been widely applied since then [22]. This model accelerated models that describe failure rate and
specifies that the effect of the load is multiplicative remaining life time of the products for a given
in time. In the AFTM, component reliability is loading history.
expressed as:
R(t ; z ) = R0 (t ⋅ φ ( z )) (20.4) 20.2.4.1 Tampered Failure Rate (TFR) Model
R 0 (.) is a survival function of an arbitrary The TFR model was first proposed by
distribution such as Weibull, Gaussian, Lognormal, Bhattacharrya & Soejoeti [29] and then generalized
and Gamma. Hence, it follows that: by Madi [30]. The acceleration of failure when the
h(t ; z ) = φ ( z ) ⋅ h0 (t ⋅ φ ( z )) stress is raised from a lower level to a higher level
(20.5) is reflected in the hazard rate function. The TFR
H (t ; z ) = H 0 (t ⋅ φ ( z ))
model is relatively simple and has been studied by
h 0 (.) and H 0 (.) are the hazard rate and cumulative several researchers [31]. In this model, the failure
hazard rate corresponding to the survival function rate of a component completely depends on the
R 0 (.). When there is only one type of load, z = L, current applied load and age of the component. In
commonly used forms of φ(L) include: other words, the failure rate is independent of the
 Tampered Failure Rate Load-Sharing Systems: Status and Perspectives 293

history of the load applied on the component. 20.3 Load-Sharing Models

However, the cumulative failure rate and reliability
of a component is a function of load history and the The majority of works on load-sharing models has
ages at which the loads are applied. As pointed out studied the failure properties of composite
in [32], it should be noted that the Khamis & materials using the concept of fiber bundles. In
Higgins model [33] is a special case of the TFR these analyses, the failure model of the individual
model where the baseline failure distribution is fibers that make up the fiber bundle is specified.
Weibull. The model can be either static or dynamic (or time-
20.2.4.2 Cumulative Exposure (CE) Model dependent). In the static case, the probability of the
failure of a fiber is specified in terms of the stress
The Cumulative Exposure model is proposed by on the fiber [12, 21]. The static models mainly
Nelson [34]. It is also called the Cumulative concentrate on the strength distribution of the
Damage model [35]. It suggests that, in calculating system (bundle of fibers) in terms of the strength
cumulative failure rate or reliability, the life-stress distribution of individual system elements (fibers).
model must take into account the cumulative effect These models generally ignore the time-varying
of the applied stresses when dealing with data from properties of materials, i.e., failure is assumed to
accelerated tests with step stresses. Alternatively, occur instantaneously. In the dynamic case the
the cumulative failure rate is calculated using an statistical distribution of times to failure for the
effective age of the component, where the effective fibers is specified in terms of stresses on the fibers
age is the sum over all load durations multiplied by [40]. Experiments generally favor the dynamic
corresponding acceleration factors. failure fiber bundle models [8, 41].
Consider a component that is subjected to the
loads z 1 , z 2 ,…, z n for the durations t 1 , t 2 ,…, t n . . 20.3.1 Static Models
Let the acceleration factor corresponding to the
load z i be α i . The effective age of the component The static models generally focus on the influence
at the baseline failure rate is then: T a = α 1 t 1 + of fiber strength, bundle length, bundle size, fiber
α 2 t 2 + …+ α n t n , whereas the actual age is T = t 1 packing, and interface properties. Most of these
+ t 2 + …+ t n . models assume that components are arranged in
The disadvantage of this model is that the parallel (parallel array of fibers) [12]. These studies
cumulative failure rate or reliability of a are also called parallel bundle theories. On the
component depends only on the duration of load, other hand, twisted bundle theories are used to
but not on the sequence of loads (no sequence study the twisted mechanical structures such as
effect) or the ages at which the loads are applied. yarns, ropes, and cables [21].
Further, the cumulative damage is additive, and the The earliest works on static models date back to
remaining life of a component depends only on the 1945. In a seminal paper, Daniels [12] showed
current stress and the current cumulative through a long and complicated proof that when
distribution function, regardless of the damage the load from fiber breaks is redistributed equally
accumulation history. Some studies suggest that it among the remaining intact fibers (equal load
is important to consider the effects of load sharing), the strength of a large bundle
sequencing on the failure behavior [36]. asymptotically approaches a Gaussian distribution.
Some variations and generalizations of these Daniels derived expressions for the asymptotic
models are presented in [37-39]. When the baseline mean and standard deviation as a function of the
failure distribution is exponential, all these models underlying strength distribution of the fibers and
coincide. In this chapter, we concentrate on the bundle size in terms of number of fibers. Here, the
TFR model while studying load-sharing systems asymptotic value for mean is independent of n.
subjected to general failure distributions. Smith [42] derived a correction factor for mean
that depends on n to improve the accuracy for
relatively small bundles. For a discussion of load-
294 Suprasad Amari, K. B. Misra, and H. Pham

sharing schemes, we refer to Harlow and Phoenix applications in reliability engineering [1-6, 58-61].
[20]. For a recent review on parallel bundle Most approximations and asympotic behaviors that
theories, we refer the reader to Phoenix and are applicable for systems with a large number of
Beyerlein [43]. elements (fiber bundles) are not applicable when
the number of components (elements) in the system
20.3.2 Time-Dependent Models is very small. Therefore, these models require a
different treatment as compared to fiber bundle
The time-dependent analysis can broadly be models. In addition, it is important to consider the
classified as: (1) finite population analysis, and (2) effects of time-varying failure rates, imperfect
asymptotic analysis. In this chapter, we concentrate switching mechanisms, repair policies, etc.
on finite population analysis where the number of In spite of a wide range of applications for load-
elements in the system is finite. Analyses for finite sharing systems, the literature on the reliability
populations are published in the reliability journals analysis of load-sharing systems is limited [2]. The
[2, 44]. On the other hand, the asymptotic behavior main reason for this is due to the inherent
(limiting behavior) is studied by assuming the complexity of these models that is introduced by
number of elements in the system approaches the load-sharing mechanisms.
infinity (are very large). These studies are As we already noted, load-sharing systems
published in physics, material science, and applied introduce dependecies among component lives.
statistics journals [40, 42-545]. Coleman [45, 46] Bivariate and multivariate distribution play
has shown that under certain conditions, as the important roles in modelling these dependencies.
number of fibers increases, the asymptotic failure However, only a certain class of multivariate
time of the bundle approaches the normal distributions is applicable for load-sharing systems
distribution. Similarly, using extreme value theory, [7]. Freund [62] first introduced a bivariate
Harlow, Smith, and Taylor [53-55] have shown exponential distribution to model a two-unit load-
that the time to failure of a parallel bundle follows sharing system.
the Weibull distribution. However, as pointed out
20.3.3.1 Freund’s Load-Sharing Model
by Borges [55], even for the case of exponential
failure time distributions, this approximation is Freund’s model is the first bivariate model that is
unsuitable from the application standpoint because physically motivated [7, 62]. Let X and Y represent
for each fixed value of t (time), the deviation from the lifetimes of components C1 and C2 of a two-
its true value increases as n (number of elements) component system. Further, assume X ~ exp(λ 1 )
increases. In addition to this, using the large and Y ~ exp(λ 2 ). According to the Freund’s model,
deviation theorem of the Cramer-Petrov type and a the failure rate of C2 changes to λ’ 2 from λ 2 (λ’ 2 >
ranking limit theorem of Loève [56], Borges λ 2 ), upon the failure event of component C1
proposed a better approximation for failure time because of extra stress. Similarly, λ 1 changes to
distribution. These results are further extended for λ’ 1 in the case where component C2 fails first, (λ’ 1
series-parallel systems [55]. Recently, Newman > λ 1 ), due to the same reason. Assuming that λ 1 +
and Phoenix [57], assuming exponential failure λ 2 - λ’ 1 ≠ 0 and λ 1 + λ 2 - λ’ 2 ≠ 0, the joint density
time distribution and a power law relationship of (X, Y) is:
between failure rate and applied load, developed
λ1λ ' 2 exp[−λ ' 2 y − (λ1 + λ 2 − λ ' 2 ) x],
asymptotic theories and new computational 
algorithms for local load-sharing models.  if 0 < x < y
f ( x, y ) = 
20.3.3 Related Models λ '1 λ 2 exp[−λ '1 x − (λ1 + λ 2 − λ '1 ) y ],

 if 0 < y < x
In this chapter, we focus on the reliability analysis
of load-sharing systems with a limited number of (20.7)
components. These models have a wide range of
 Tampered Failure Rate Load-Sharing Systems: Status and Perspectives 295

It should be noted that the marginal failure, the unfailed component carries the full
distributions of X and Y are not exponential load. Hence, we have:
distributions except for the special cases: λ’ 1 = λ 1
λ1 + λ 2
and λ’ 2 = λ 2 . However, the marginal distributions R(t ) = e −( λ + λ
1 2 )t
+
λ1 + λ 2 − λ '
(
e −( λ + λ )t − e −λ 't1 2
)
can be shown to be mixtures or weighted averages
of exponential distributions [7]. (20.11)
If the system functions as long as 1-out-of-2 If the two components are equal and the load is
components are functioning, the system failure distributed equally among these two components,
time is equal to T = max(x, y). Integrating the we have: λ 1 = λ 2 = λ , and λ’ 1 = λ’ 2 = λ’. Hence,
density function over the region t < T gives the we have:
unreliability of the system. Therefore, the system
unreliability is: R(t ) = e − 2 λt +
2λ
2λ − λ '
(
e − 2 λt − e −λ 't ) (20.12)
λ1  λ ' 2 ⋅e − ( λ1 + λ2 ) t
− (λ1 + λ 2 ) ⋅ e − λ '2 t

Q(t ) = 1 +  Only a few reliability engineering textbooks that
λ1 + λ 2  λ1 + λ 2 − λ ' 2 
 discuss the load-sharing systems cover more than
 λ '1 ⋅e −( λ + λ )t − (λ1 + λ 2 ) ⋅ e −λ ' t  this simple model [1, 6, 59].
λ2 1 2 2

+ 1 +  Freund’s model can be viewed as a simple load-

λ1 + λ 2  λ1 + λ 2 − λ '1 
 sharing model for a system with two components.
(20.8) As noted above, Freund did not consider the
Therefore, the system reliability is: underlying load-sharing rules that dictate how
failure rates change after some components in the
λ1  (λ1 + λ 2 ) ⋅ e − λ ' t − λ ' 2 ⋅e − ( λ + λ
2 1 2 )t
 system fail. Weier [66] is the first one who actually
R(t ) =  
λ1 + λ 2  λ1 + λ 2 − λ ' 2 
 analyzed the reparametrization of the Freund’s
model. He modelled the post failure hazard rate θ 2
λ2  (λ1 + λ 2 ) ⋅ e −λ ' t − λ '1 ⋅e −( λ + λ
 2 )t
by γθ 1 , γ > 0. Here γ=1 implies independence, and
2 1

+  
λ1 + λ 2  λ1 + λ 2 − λ '1
 γ>1 corresponds to an increased work load on the
(20.9) remaining component, while γ<1 corresponds to a
The reliability expression can be rearranged as: reduced work load. This parameterization allows
the researchers to extend the Freund’s model for
 λ1 λ2  general cases, such as k-out-of-n systems, and
R(t ) = e −( λ + λ )t 1 −
1 2
− 

 λ1 + λ 2 − λ ' 2 λ1 + λ 2 − λ '1  make general statistical inferences on the details of
possible dependencies among components in a
λ1 λ2
+ e −λ ' t +2
e −λ ' t 1
system [7] .
λ1 + λ 2 − λ ' 2 λ1 + λ 2 − λ '1
20.3.3.2 k-out-of-n System with IID Components
(20.10)
The above system behavior can be modelled using Scheuer [44] studied the reliability of a k-out-of-
a four-state Markov chain, and the closed-form n:G system where component failure induces
solution can be obtained using either convolution higher failure rates in the survivors. His work
integrals or Laplace transformations [63, 64]. The assumed that the components are IID with constant
above model is applicable for all types of load- failure rates. Although it is not mentioned
sharing rules: equal, local, or monotone load- explicitly, the model inherently assumes the equal
sharing rules. In fact, this model is independent of load-sharing rule. Scheuer modelled the system
the load-sharing rule. failure time as the sum of independent exponential
Now assume that λ 1 ≠ λ 2, but λ’ 1 = λ’ 2 = λ’. distributions. He identified three cases.
This assumption is valid if both the components are • Case-1 arises when all these exponential
identical and load is distributed unequally when distributions are identical. In this case, the
both components are working. However, after a
296 Suprasad Amari, K. B. Misra, and H. Pham

system failure follows the Erlang distribution 20.3.3.4 Multivariate Exponential Models
(a special case of the Gamma distribution).
Lin, Chen, and Wang [74] extended the Freund’s
• Case-2 arises when all these exponential bivariate exponential model to the multivariate
distributions are distinct. In this case, the exponential case. In this model, the system consists
system failure time follows the of n non-identical components. The failure rate of a
Hypoexponential distribution. component depends on the number of failed
• Case-3 arises when all these exponential components, but it is independent of the actual set
distributions are neither equal nor distinct. of components that are failed. For example, when
This happens when distinct groups of there is only one failed component in the system,
distributions exist. Within each of these the failure rate of component 1 is the same,
groups, all of the distributions are equal. In regardless as to whether component 2 or
this case, the system failure time is the sum component 3 failed. This assumption is applicable
of Erlangian distributions with different for the equal load-sharing rule, where the load on a
parameters. Scheuer [44] mentioned that surviving component is a function of the number of
there is no convenient closed-form solution failed components. Even with this assumption, the
for this case. However, recently Amari and failure time distribution of a 1-out-of-3 system is
Misra [65] provided a closed-form solution to too complex, and the equation for the system
this case. This case has a wide range of reliability occupies almost two full columns of that
applications not only in reliability paper. Lin et al [74] also presented a closed-form
engineering but also in other fields such as solution for k-out-of-n systems with IID
control systems and telecommunications [67, components following exponential failure
68]. distributions. However, this is equivalent to the
case-2 studied by Scheuer [44], where the closed-
20.3.3.3 k-out-of-n System with Repair
form solution is well known. Later on, Amari [64]
Shao and Lamberson [69] provide an analysis of a provided a compact closed-form solution to k-out-
repairable k-out-of-n:G system with load-sharing of-n systems with non-identical components.
components considering imperfect switching. In
20.3.3.5 General Failure Distributions
this model, all components in the systems are
identical and follow exponential failure There is not much published work on the reliability
distributions. As in most cases, it also assumes the analysis of load-sharing systems with components
equal-load sharing rule. The sensing and switching subjected to general failure distributions. Even
mechanism is responsible for detection of though there are some papers, they do not
component failures and the redistribution of the explicitly mention the underlying assumptions. In
load of the system equally among surviving modelling load-sharing systems with general
components. System performance measures such distributions, it is important to consider an
as reliability and availability are analyzed using appropriate model to incorporate the effects of
Markov chains. Unfortunately, several errors exist loading history. For details, refer to section 20.2.4.
in this paper, which are corrected by Akhtar [70]. Several researchers have extended the Freund’s
Newton [71] provides an alternative argument for bivariate exponential model to other cases that
evaluation of the MTTF and MTBF of such include the case of non-exponential distributions
systems. A corrected version of this model is such as Weibull and Gamma [75, 76]. However,
presented in [72]. However, [72] did not provide a only a few of these extensions are useful for
complete solution to this model, but provided the modelling load-sharing systems. Lu [76] proposed
differential equations for solving the Markov a bivariate Weibull distribution that can be applied
chains. Recently, Chan [73] discussed availability to load-sharing systems. In this model, the failure
analysis of a load-sharing system and advocated distribution (parameters) of the surviving
the combining of performance and reliability component changes after load redistribution due to
analyses within the Markov reward framework. the failure of other component. However, in this
 Tampered Failure Rate Load-Sharing Systems: Status and Perspectives 297

model, it is assumed that the effective age of the General Assumptions

component becomes zero after the load
redistribution, which is not a realistic assumption. 1. After a component failure, the load is equally
Kececioglu and Jiang [77] presented a 1-out- distributed among all surviving components.
of-2 load-sharing system with Weibull failure time This assumption will be relaxed in some
distributions. In this analysis, it is assumed that cases.
after load redistribution, the component age 2. The failure rate of a component varies as per
changes according to an accelerated failure-time the TFR model. The baseline failure rate of
model (AFTM) as per the cumulative exposure the TFR model can follow an arbitrary
model. This model is also available in [59]. Later distribution such as Weibull, Gaussian,
Liu et al [26, 78, and 79] proposed similar concepts Lognormal, and Gamma.
for analyzing 1-out-of-2 parallel systems subjected 3. The redistribution and reconfiguration
to variable working conditions. Liu [2] further mechanisms are perfect.
extended this model for analyzing load-sharing k- 4. The system and its components are non-
out-of-n:G systems with arbitrary load-dependent repairable.
component lifetime distributions. However, the
solution provided in [2] is too complex. Therefore, 20.4.1 Load Distribution
as mentioned in [2] and [6], the solution can only
be applied for simple systems where n ≤ 6 (even In a load-sharing system, upon a component
using MatLab). Therefore, more efficient methods failure, the load on the failed component is
for handling arbitrary load-dependent component redistributed among the surviving components. In a
lifetime distributions are needed. majority of cases, the load is equally distributed
Recently, Amari, Misra, and Pham [19] over all surviving components. If the total load is
provided a closed-form analytical solution for the L, and there are m good components, then the load
reliability of tampered failure rate (TFR) load- on each component is z = L/m.
sharing k-out-of-n:G systems with identical Let n be the total number of components in the
components where all surviving components share system and z i be the load on each of the surviving
the load equally. This model can be considered as a components when i components are failed. Hence,
generalization of the 1-out-of-2 load-sharing model n
presented in Hassett [80], where the solution is z 0 = L / n; z i = L /(n − i ) = z 0 ⋅ (20.13)
n−i
provided using a numerical integration technique
applicable for semi-Markov chains [81]. 20.4.2 The TFR Model

20.4 System Description In the TFR model, the acceleration of failure when
the stress is raised from a lower level to a higher
In this chapter, we provide a detailed treatment of level is reflected in the hazard rate function. In this
TFR load-sharing systems subjected to exponential section, we describe this model in terms of k-out-
and general failure time distributions. Because k- of-n systems. Consider a component that is
out-of-n systems have a wide range of applications subjected to an ordered sequence of loads, where
in reliability engineering, a detailed analysis of load z i (i=0,1,…,n-k) is applied during the time
these systems shall be presented in this chapter. interval [t i , t i+1 ], where t 0 = 0. In other words, the
The concepts presented here can easily be extended load changes at times t 1 , t 2 ,…, t n-k . According to
to analyze general system configurations. In this the TFR model, the hazard rate of the component at
chapter, we also would like to relax the IID time t is:
assumption used in [19]. h(t ) = hi (t ) = δ i ⋅ h0 (t ) for t i −1 ≤ t < t i (20.14)
where δ 0 =1, h 0 (t) is the hazard rate at the lower
load z 0 , and δ i is the tampered factor at load level
298 Suprasad Amari, K. B. Misra, and H. Pham

z i . The tampered factor is a function of the applied i th failure occurs at rate α i = (n-i+1).λ i-1 . The
stress. Hence, the TFR model can be expressed as: system is failed when more than (n-k) components
h(t ) = δ ( z ) ⋅ h0 (t ) (20.15) are failed.
where z is the load at time t. Because all components are IID following the
exponential distributions, the inter-arrival times of
20.4.3 System Configuration failures are independent random variables X i ,
where X i follows the exponential distribution with
In this chapter, we considered a k-out-of-n parameter α i for 1≤ i ≤ n-k+1. Hence, the lifetime
structure, which is a common form of redundancy of the system is equal to the (n-k+1)st failure time.
used in reliability engineering. A system is called Alternatively, the lifetime of the system is equal to
called a k-out-of-n system if at least k out of n the sum of (n-k+1) independent random variables
components must work for the successful operation following exponential distributions with possibly
of the system [1, 6]. The k-out-of-n:G structure different parameters (rates).
redundancy finds wide applications in both
industrial and military systems. Several examples T = X 1 + X 2 + ... + X n − k +1
(20.16)
of k-out-of-n:G systems are available in [1, 6]. Hence R(t ) = Pr{T ≤ t}
Both series systems and parallel systems are
special cases of the k-out-of-n system. To find the distribution of T and the reliability
function of the system, we need to distinguish the
following three cases.
20.5 k-out-of-n Systems with Identical
Components Case-1: All α I ’s are equal (say α) [44].
n−k
(at ) i exp(−at )
In this section, we discuss the load-sharing k-out- R(t ) = ∑i =0 i! (20.17)
of-n systems with identical components. Therefore,
the additional assumptions are: = gamfc(at ; n − k + 1)
• There are n IID components in the system. where gamfc() is the complimentary cumulative
• The system functions successfully if and only distribution of the Gamma distribution. This case
if there are at least k good components. arises when the failure rate of each surviving
component is directly proportional to the load it
20.5.1 Exponential Distributions carries. Hence, in the TFR model, δ(z) ∝ z.

When the system is put into operation at time zero, Case-2: All α I ’s are distinct [44].
all components are working, and they are equally n − k +1
sharing the constant load that the system is R(t ) = ∑ A ⋅ exp(−α t )
i =1
i i
supposed to carry. In this case, the failure rate of (20.18)
every component is denoted by λ 0 . Because there n − k +1 αj
are n working components in the system, the first
Ai = ∏
j =1; j ≠ i α i + α j
failure occurs at rate α 1 = n .λ 0 . When the system
experiences the first failure, the remaining (n-1)
working components must carry the same load on Case-3: α I ’s are neither equal nor distinct.
the system. As a result, the failure rate of each Specifically, assume that these α I ’s take a (1 < a <
working component becomes λ 1 , which is typically n-k+1) distinct values, β 1 , β 2 ,…, β a . With possibly
higher than λ 0 . The second failure occurs at rate α 2 some renumbering of these α i ’s values, assume:
= (n-1).λ 1 . When i components are failed, the
failure rate of each of the (n-i) working
components is represented by λ i (0 ≤ i ≤ n-k). The
 Tampered Failure Rate Load-Sharing Systems: Status and Perspectives 299

a 1 =  = a r = β1 1
when there are some additional spares in standby.
In such cases, the load on the surviving
a r +1 =  = a r + r = β 2
1 2 2 components remains the same until the exhaustion
(20.19)
 of the standby spares. Hence, the values of some
a r + r ++ r =  = a r + r ++ r = β a α I ’s will be the same.
2 2 a −1 +1 2 2 a

20.5.2 General Distributions

where the assumptions are:
• a ≥ 1 and an integer. In this section, we show that solving TFR models
• All β I ’s i are distinct. with arbitrary baseline failure distributions is
• Sum of r I ’s is equal to n. equivalent to solving TFR models with exponential
distributions. The basic idea is that using a time-
• Each r i ≥ 1 and an integer.
transformation, a TFR model with an arbitrary
Hence, T = V1 + V 2 + … + V a . V i is a random
baseline distribution can be converted into an
variable, which follows the Erlangian distribution.
equivalent problem with an exponential baseline
The shape parameter is r i , and the scale parameter
distribution. This in turn reduces the problem under
is β i . From [44], we have:
consideration to a simplified problem: the problem
a rj Φ jl (- β j )
R(t ) = B ∑∑
j =1 l =1 (l - 1)!( β j )
r j -l +1
.poif (r j -l;β j ⋅ t ) of a k-out-of-n:G load-sharing system with
exponential distributions.
where
a
Lemma 1: For any failure distribution F(t):
B = ∏ (β
rj
j ) 1. Reliability function is R(t) = 1-F(t).
j =1
2. Cumulative hazard function is H(t) =
 a  ln[R(t)].
Φ jl (t ) = D l −1 
 i =1;i ≠ j∏
(β i + t ) −r 

(20.20)
i
3. The function H(t) is a non-decreasing
  function in t.
r −l
exp(− β j t )( β j t ) i 4. The random variable y = H(t) follows an
∑
j

poif (r j − 1; β j ⋅ t ) ≡ exponential distribution with a mean of 1.

i =0 i!
5. For any constant ‘a’, the random variable y =
a .H(t) follows an exponential distribution
Scheuer [44] mentioned that there seems to be no
with a mean of ‘1/a’ and a failure rate of ‘a’.
closed-form expression for Φ jl (t). Using a multi-
function generalization of the Leibnitz rule, Amari Proof: #1-#3 are straightforward, basic reliability
and Misra [65] provided a closed-form solution for engineering concepts. Because R(t) = e-y, #4-#5
Φ jl (t), which in turn completes the closed-form can easily be proved using transformation of
solution for R(t). variables (see [1]).
a
Φ jl (t ) = (−1) l (l − 1)! ∑ ∏ν (i, m , t )
Ω i =1;i ≠ j
i Lemma 2: For a TFR model with a standard
exponential (mean = rate = 1) baseline failure time
 ri + mi − 1 distribution:
ν (i, mi , t ) =  ( β j + t ) −( r + m ) i i

 m i  h(t ) = δ i
Ω ≡ m1 +  + ma = l − 1; where m j = 0; mi ≠ j ≥ 0 H (t ) = H (t i −1 ) + δ (t − t i −1 ) (20.22)
(20.21) for t i −1 ≤ t < t i
The Case-3 arises when δ(z) in the TFR model is
either a piece-wise linear function or a S-shared Proof: Straightforward from (20.15).
increasing function. The Case-3 is also applicable
300 Suprasad Amari, K. B. Misra, and H. Pham

Lemma 3: For a TFR model with a baseline failure system can be calculated using λ i = δ i , where λ i is
rate of h 0 (t) and a baseline cumulative failure rate the failure rate of a component when i components
of H 0 (t): are failed.
1. Under the regular scale t: As long as the baseline failure rate in the TFR
model is the same, theorem 1 is also applicable for
h(t ) = δ i ⋅ h0 (t )
non-identical component cases where δ i ‘s are
H (t ) = H (t i −1 ) + δ i ⋅ [H 0 (t ) − H 0 (t i −1 )] (20.23) different for the non-identical components. It
for t i −1 ≤ t < t i should be noted that the results of theorem 1 are
not restricted to k-out-of-n:G systems. However, in
2. Under the transformed scale y ≡ H 0 (t): this chapter, we apply these results for solving k-
out-of-n:G systems.
Leτ ν i = H 0 (τ i )
h y ( y) = δ i 20.5.3 Examples
(20.24)
H y ( y ) = H y (ν i −1 ) + δ i ⋅ [ y − ν i −1 ]
Example 1: Consider a 5-out-of-10:G system with
for ν i −1 ≤ y < ν i Weibull as the baseline failure distribution.
where H y (y) is the cumulative hazard rate in the
transformed scale. Model : k = 5; n = 10; t = 1000
Proof: #1 is straightforward from the definition of β −1 β
cumulative hazard rate and the h(t) in (20.15). β t t
Baseline : h0(t) = ⋅   ; H 0(t) =  
Because t = H 0 -1(y) and H 0 (t) = y, #2 follows η η η
from #1.
  t β 
R0 (t ) = exp[− H 0 (t )] = exp −   
Theorem 1: If the effects of load variations on the   η  
hazard rate of an individual component follow a
TFR model with h(t) = δ i .h 0 (t), the reliability of a Baseline Parameters : η = 2000; β = 2
load-sharing system at time t is equivalent to the TFR : h(t) = δ(z) ⋅ h0(t)
reliability of the corresponding exponential load- 1.5
 n 
sharing model at time y = H 0 (t), where the failure δ ( z) = z ⇒ δ i = δ ( zi ) = 
1.5

rate of a component when i components failed is λ i n−i
= δ i for i = 0,…, (n-k). Soλution : y = H 0 (t ) = 0.25; λi = δ i
α i = (n − i ) ⋅ λi = 31.623(n − i ) −0.5
Proof: There is a one-to-one relationship between
reliability and cumulative failure rate. For example, Hence : α 1 = 10; α 2 = 10.541; α 1 = 11.18;
if the cumulative failure rates of two components at α 4 = 11.952; α 5 = 12.91; α 6 = 14.142
two different time points are equal, then their Here all α i 's are distinct. Hence, from (20.18),
reliabilities are also equal at those points. R y (0.25) = 0.92415 ⇒ R(t=1000)= 0.92415.
Mathematically, if H 1 (t 1 ) = H 2 (t 2 ), then R 1 (t 1 ) =
R 2 (t 2 ). From lemma 2 & lemma 3, H y (y) = H(t),
Example 2: Same as example 1, except: δ(z) = z.
where Hy(y) is the cumulative failure rate of a
component with an arbitrary baseline failure rate in Solution : δ (z) = z ⇒ α 1 =  = α 6 = α = 10
the transformed scale y, and H(t) is the cumulative Here all α i 's are equal. Hence, from (20.17),
failure rate of a component with a constant baseline R y (0.25) = 0.95798 ⇒ R(t=1000)= 0.95798.
failure rate (rate = 1) in the regular scale.
Therefore, under the transformed scale, all TFR
Example 3: Same as example 1, except: δ(z) is
models are equivalent to their corresponding
piece-wise linear in z.
constant failure rate models. Hence, under the
transformed scale, the reliability of a load-sharing
 Tampered Failure Rate Load-Sharing Systems: Status and Perspectives 301

 z : z < 1.2 even for the 1-out-of-3 systems. In this section, we

 provide a different approach for solving these
TFR : δ(z) = 1.1z : 1.2 ≤ z < 1.5
systems.
 1.2 z : z ≥ 1.5
 One approach to solve this problem is
Solution : α 1 = α 2 = 10; α 3 = α 4 = 11; α 5 = α 6 = 12 automatic generation of Markov chains. In this
approach, as in dynamic fault tree analysis [82, 83],
Here α i 's are neither equal nor distinct. Hence,
we generate Markov chain that describe the system
from (20.20, 20.21), R y (0.25) = 0.94005 ⇒
behavior and solve the Markov chain using
R(t=1000)= 0.94005.
numerical integration methods. The number of
Example 4: Same as example 2: δ(z) = z. states generated is:
Additionally, it has two spares in standby, which n − k +1
n
are used to replace the first two failed components. N = ∑  
i =0  i 
(20.25)
The failure rate of a spare in standby is:
If we merge all failed states into one single state,
In Standby : h(t ) = γ ⋅ h0 (t ); where γ = 0.5 then we have:
Solution : α 1 = 11, α 2 = 10.5; α 3 = 3 = α 8 = 10
n−k
n
Here α i 's are neither equal nor distinct. Hence,
∑
N = 1 +  
i =0  i 
(20.26)

from (20.20, 20.21), R y (0.25) = 0.98433 ⇒ The number of states increases with both n and n-k.
R(t=1000)= 0.98433. For a parallel system, the number of states is equal
Example 5: Same as example 4, except the spare to 2n. It should be noted that if the failure rate is
are in cold standby where the failure rate is zero, dependent on both the number of component
and δ(z) = z1.5. failures and the actual set of failed components,
Solution : α 1 = α 2 = α 3 = 10; α 4 = 10.541; then the number of states increases drastically. In
the later case, because a failed state with k
α 5 = 11.18; α 6 = 11.952; α 7 = 12.91; α 8 = 14.14
component failures can be reached in k! ways, a
Here α i 's are neither equal nor distinct. Hence, parallel system with 10 components can have as
from (20.20, 20.21), R y (0.25) = 0.97874 ⇒ many as 9,864,101 states [83]. The existing
R(t=1000)= 0.97874. Markov chain solvers can handle a maximum of
100,000 states. Hence, 2n ≤ 100,000. Therefore, n
should be less than or equal to 16. There are
20.6 k-out-of-n Systems with Non-
several approaches to improve the efficiency of the
Identical Components computation. They include: (1) eliminating the
states that contribute an insignificant portion of
In this section, we consider a k-out-of-n system probability to the failure, (2) solving the Markov
with non-identical components. We first discuss chains while generating the states, and (3) utilizing
the exponential distribution case that is considered the bounds. Obviously, another alternative is to use
in Lin, Chen, and Wang [74] and later extend this simulation methodology. However, in all these
model to the general failure distribution case. methods, we can only have numerical solutions.
In some cases, we may be interested in the
20.6.1 Exponential Distributions closed-form analytical solutions. We use a method
proposed in [64] to obtain the closed-form
In this model, the system consists of n non- solutions to this problem. The method is
identical components. The failure rate of demonstrated using a 2-out-of-3 load-sharing
component i when there are j failures in the system system. In this method, we first generate all
is λ(i,j). Because the system reaches a failed state sequences of component failures that lead to the
when there are n-k+1 failed components, we need system failure. The sequences are mutually
to consider the j values in the region: 0 ≤ j ≤ n-k. exclusive. In this case, we have 6 failure
The existing analysis for this model is too complex sequences. They are: {1, 2}, {1,3}, {2,1}, {2,3},
302 Suprasad Amari, K. B. Misra, and H. Pham

{3,1} and {3,2}. The overall system failure all these components is the same. (or failure
probability is the sum of the probability of each rates/are) Hence, we have:
failure sequence. The Markov model for this h (i, j , t ) = λ (i, j ) ⋅ h0 (t ) (20.30)
sequence is shown in Figure 20.1. The labels in the Therefore, extending the concepts used for the
Markov chain explain the properties of the states indentical component case, we can compute the
and transitions. system reliability with general failure distributions
using the solutions that are applicable for the
All Good 1st Failure 2nd Failure exponential failure distributions.
a 1 = l (2,0) a 2 = l (3,1)
{} {2} {2,3}
Good Good Failed
Theorem 2: If the effects of load variations on the
hazard rate of an individual component follow a
TFR model with h(i,j,t) = λ(i,j).h 0 (t), the reliability
β1 = l (1,0) + l (3,0) β 2 = l (1,1) of a load-sharing system at time t is equivalent to
the reliability of the corresponding exponential
Others load-sharing model at time y = H 0 (t), where the
Seq
failure rate of a component i when j components
are failed is λ(i,j).
Fig. 20.1: Markov Chain for Sequence {2,3}
Proof: This is similar to the proof of Theorem 1.
The Laplace transform for the failed state is:
20.6.3 Further Examples
α 1α 2
F (s) = (20.27)
s ( s + α 1 + β1 )( s + α 2 + β 2 ) Example 6: Consider a 2-out-of-3:G system with
Weibull as the baseline failure distribution.
Using [65], the time-dependent failure probability
contribution of sequence {2, 3} is: Model : k = 2; n = 3; t = 1000
β −1 β
F (t ) = A0 + A1 exp(−γ 1t ) + A2 exp(−γ 2 t ) β t t
Baseline : h0(t) = ⋅  
; H 0(t) =  
(20.28) η η η
where
  t β 
γi = αi + βi R0 (t ) = exp[− H 0 (t )] = exp −   
2
αi   η  
A0 = ∏γ
i =1 i
(20.29)
Baseline Parameters : η = 2000; β = 2
αj 2
αi TFR : h(i, j , t) = λ (i, j ) ⋅ h0(t)
A j ≠0 = −
γj ∏γ
i =1;i ≠ j i −γ j λ (i, j ) = δ i ( z j )
where n-k+1 = 2. For the general case, we can n n
zj = ⋅ z0 = (where z 0 = 1)
replace the 2 with n-k+1. n−i n−i
Hence:
20.6.2 General Distributions bi
 n 
δ i ( z j ) = ai ⋅ ( z i ) b ⇒ λ (i, j ) = ai ⋅ 
i

In this model, the system consists of n non- n−i
identical components. The failure rate of a1 = 1; a 2 = 1.2; a 3 = 1.3;
component i at time t when there are j failed b1 = 1; b2 = 1.1; b3 = 1.2;
components in the system is h(i,j,t). In addition to
this, we also assume that the baseline failure rate of
Hence:
 Tampered Failure Rate Load-Sharing Systems: Status and Perspectives 303

λ (1,0) = 1.0; λ (2,0) = 1.2 λ (3,0) = 1.3; explosion of the Markov chains, the semi-Markov
λ (1,1) = 1.5; λ (2,1) = 1.875 λ (3,1) = 2.115; solution approach cannot directly be applied to
solving large models. Hence, there is a need for
Solution : y = H 0 (t ) = 0.25; better solutions. Similarly, at present, there are no
Sequence probabilities at y = 0.25 are: efficient methods for solving load-sharing
cumulative exposure models. In addition to this, it
F (1,2) = 0.032195; F (1,3) = 0.036321 is interesting to study the optimal dynamic load
F (2,1) = 0.031808; F (2,3) = 0.044842 distribution of these systems.
F (3,1) = 0.035101; F (3,2) = 0.043865
Hence, the overall failure probability is 0.224131. References
Therefore, from Theorem 2, R y (0.25) = 0.775869
⇒ R(t=1000)= 0.775869. [1] Misra KB. Reliability analysis and prediction: A
It should be noted that as long as the baseline methodology oriented treatment. Elsevier, 1992.
[2] Liu H. Reliability of a load-sharing k-out-of-n:G
failure rate is the same, this procedure can also be
system: non-iid components with arbitrary
used to compute the reliability of non-equal load- distributions. IEEE Trans. on Reliability 1998; 47:
sharing rules. 279-284.
[3] Kapur KC, Lamberson LR. Reliability in
engineering design. John Wiley & Sons, 1977;
20.7 Conclusions 405-414.
[4] Iyer RK, Rossetti DJ. Effect of system workload
In this chapter, various concepts that are important on operating system reliability: a study on IBM
in the modelling and analysis of load-sharing 3081. IEEE Trans. on Software Engineering 1985;
systems, have been presented along with a state-of- SE-11:1438-1448.
the-art review on existing modelling techniques [5] Iyer RK, Rosetti DP. A measurement-based model
for workload dependency of CPU errors. IEEE
and solution methods. Although load-sharing
Trans. on Computers 1986; C-35: 511-519.
systems have a wide range of applications, due to [6] Kuo W, Zuo MJ. Optimal reliability modelling.
the inherent complexity of these systems, the Wiley, 2003; 258-264.
methods available for analyzing these systems are [7] Kim H. Reliability modelling with load-shared
limited. data and product-ordering decisions considering
The existing literature covers only a small uncertainty in logistics operations. PhD
portion of these systems that include exponential Dissertation, Georgia Institute of Technology
failure time distributions, equal load-sharing rules, 2004.
and identically distributed components. Therefore, [8] Kvam PH. Peña EA. Estimating load-sharing
properties in a dynamic reliability system. J.
in order to widen state of the art, we have proposed
Amer. Statistical Assoc. 2005; 100: 262-272.
efficient methods for computing the reliability of [9] Aven T, Jensen U. Stochastic models in reliability.
load-sharing k-out-of-n:G systems with identical Springer, 1999.
and non-identical components where all surviving [10] Nakagawa T. Shock and damage models in
components share the load equally. The method reliability theory. Springer, 2006.
can be applied for a wide range of failure time [11] Marshall AW, Olkin I. A multivariate exponential
distributions including Weibull, Lognormal, and distribution. J. Amer. Stat. Assoc. 1967; 62: 30-44.
Gamma distributions. As long as the baseline [12] Daniels H. E. The statistical theory of the strength
failure time distribution of all components are the bundles of threads I. Proc. Royal Society of
London, Series A 1945; 183: 405–435.
same, the solution proposed in this chapter is also
[13] Carlson RL, Kardomateas GA. An introduction to
applicable for unequal load-sharing rules. It may be fatigue in metals and composites. Chapman &
observed here that all TFR models including non- Hall, 1996.
identical baseline failure rate cases can also be [14] Jelinski Z, Moranda P. Software reliability
solved using semi-Markov chains [64, 80]. research. In: Freiberg W, editor. Statistical
However, due to the well-known state space
304 Suprasad Amari, K. B. Misra, and H. Pham

computer performance evaluation. Academic test. Communications in Statistics: Theory Method

Press, 1972; 465–484. 1989; 18: 1627–1643.
[15] Wang YT, Morris RJT. Load sharing in distributed [30] Madi M. T., Multiple step-stress accelerated life
systems. IEEE Trans. on Computers 1985; 34: test: The tampered failure rate model.
204-217. Communications in Statistics, Theory and
[16] Kvam P. H. and Day D., The multivariate Polya Methods 1993; 22: 2631–2639.
distribution in combat modeling. Naval Research [31] Wang R. and Fei H., Uniqueness of the maximum
Logistics 2001; 48: 1–17. likelihood estimate of the Weibull distribution
[17] Jewell N. P. and Kalbfleisch J. D., Marker tampered failure rate model. Communications in
processes in survival analysis. Lifetime Data Statistics: Theory and Methods 2005; 32: 2321-
Analysis 1996; 2: 15–19. 2338.
[18] Amari S. V., McLaughlin L., and Pham H., Cost- [32] Xu H. and Tang Y., Commentary: The
effective condition-based maintenance using Khamis/Higgins model. IEEE Trans. on Reliability
Markov decision processes. Proc. Annual 2003; 52: 4-6.
Reliability and Maintainability Symp. 2006; 464- [33] Khamis H. and Higgins J. J., New model for step-
469. stress testing. IEEE Trans. on Reliability 1998; 47:
[19] Amari S. V., Misra K. B., and Pham H., Reliability 131-134.
analysis of tampered failure rate load-sharing k- [34] Nelson W., Accelerated life testing-step-stress
out-of-n:G systems. Proc. 12th ISSAT Int'l Conf. models and data analysis. IEEE Trans. on
on Reliability and Quality in Design 2006; pp. 30- Reliability 1980; R-29: 103–108.
35. [35] Mettas A. and Vassiliou P., Application of
[20] Harlow D. and Phoenix S. L, The chain-of-bundles quantitative accelerated life models on load
probability model for the strength of fibrous sharing redundancy. Proc. Ann. Reliability and
materials. I - Analysis and conjectures. Journal of Maintainability Symp. 2003; 551-555.
Composite Materials 1978; 12: 195-214. [36] Van Paepegem W. and Degrieck J., Effects of
[21] Porwal P. K., Beyerlein I. J., and Phoenix S. L., Load Sequence and Block Loading on the Fatigue
Statistical strength of a twisted fiber bundle: an Response of Fiber-Reinforced Composites. Mech.
extension of Daniels equal-load-sharing parallel of Adv. Materials and Structures 2002; 9: 19-35.
bundle theory. Journal of Mechanics of Materials [37] Degroot M. H. and Goel P. K., Bayesian
and Structures 2006; 1: 1425-1447. estimation and optimal design in partially
[22] Nelson W., Accelerated Testing: Statistical accelerated life-testing. Nav. Res. Logist. Quart.
Models, Test Plans, and Data Analysts. John 1979; 26: 223–235.
Wiley & Sons 1990. [38] Zhao W. and Elsayed E., A general accelerated life
[23] Cox D. R., Regression models and life tables (with model for step-stress testing. IIE Transactions
discussion). J. Royal Stat. Soc., Series B 1972, 2005; 37: 1059-1069.
34(2): 187-220. [39] Pan R. and Ayala S., Two statistical models for
[24] Jardine A. K. S., Ralston P., Reid N., and Stafford step-stress accelerated life test analysis. 47th Fall
J., Proportional hazards analysis of diesel engine Tech. Conf. of ASQ and ASA 2003.
failure data. Quality and Reliability Engineering [40] Coleman B. D., Statistics and time dependence of
International 1989; 5: 207-216. mechanical breakdown in fibers. Journal of
[25] Kumar D. and Klefsjo B., Proportional hazards Applied Physics 1958; 29: 968–983.
model: A review. Reliability Engineering and [41] Rundle J. B., Turcotte D. L., Shcherbakov R.,
System Safety 1994; 44: 177-188. Klein W., and Sammis C., Statistical physics
[26] Liu H. and Makis V., Cutting-tool reliability approach to understanding the multiscale
assessment in variable machining conditions. dynamics of earthquake fault systems. Rev.
IEEE Trans. on Reliability 1996; 45: 573-581. Geophys 2003; 41: 5.1-5.30.
[27] Pike M. C., A method of analysis of a certain class [42] Smith R. L., The asymptotic distribution of the
of experiments in carcinogenesis. Biometrics strength of a series-parallel system with equal
1966; 22: 142-161. load-sharing. The Annals of Probability 1982; 10:
[28] Solomon P. J., Effect of misspecification of 137-171.
regression models in the analysis of survival data. [43] Phoenix S. L. and Beyerlein I. J., Statistical
Biometrika 1984; 71: 291-298. strength theory for fibrous composite materials,
[29] Bhattacharyya G. K. and Soejoeti Z., A tampered Comprehensive composite materials, Editor: Kelly
failure rate model for step-stress accelerated life A., et al, Elsevier 2000; 1: 559–639.
 Tampered Failure Rate Load-Sharing Systems: Status and Perspectives 305

[44] Scheuer E. M., Reliability of an m-out-of-n system [59] Kececioglu D., Reliability Engineering Handbook.
when component failure induces higher failure PTR Prentice Hall 1991; 2: pp. 363-399.
rates in survivors. IEEE Trans. on Reliability [60] Pham H., Reliability analysis of a high-voltage
1988; 37: 73-74. power system with dependence failure and
[45] Coleman B. D., Time Dependence of Mechanical imperfect coverage. Reliability Engineering and
Breakdown in Bundles of Fibers I: Constant Total System Safety 1992; 37: 25-28.
Load. Journal of Applied Physics 1957; 28: 1058- [61] Birolini A., Reliability Engineering: Theory and
1064. Practice. Springer-Verlag 2004.
[46] Coleman B. D., Time Dependence of Mechanical [62] Freund J. E., A bivariate extension of the
Breakdown in Bundles of Fibers II: The Infinite exponential distribution. J. of Amer. Stat. Assoc.
Ideal Bundle under Linearly Increasing Loads. 1961; 56: 971–977.
Journal of Applied Physics 1957; 28: 1065-1067. [63] Pozsgai P., Neher W., and Bertsche B., Models to
[47] Kuo C. C. and Phoenix S. L., Recursions and limit consider load-sharing in reliability calculation and
theorems for the strength and lifetime distributions simulation of systems consisting of mechanical
of a fibrous composite; Journal of Applied components. Proc. Ann. Reliability and
Probability 1987; 24: 137-159. Maintainability Symp. 2003; 493-499.
[48] Phoenix S. L., The asymptotic distribution for the [64] Amari S. V., Reliability, Risk and Fault-Tolerance
time to failure of a fiber bundle. Advances in of Complex Systems. PhD Dissertation, Indian
Applied Probability 1979; 11: 153-187. Institute of Technology, Kharagpur 1997.
[49] Phoenix S. L., The asymptotic time to failure of a [65] Amari S. V. and Misra R. B., Closed-form
mechanical system of parallel members. SIAM expressions for distribution of sum of exponential
Journal on Applied Mathematics 1978; 34: 227- random variables. IEEE Trans. on Reliability
246. 1997; 46: 519-522.
[50] Phoenix S. L., Probabilistic theories of time [66] Weier D. R., Bayes estimation for bivariate
dependent failure of fiber bundles. Oceans 1976; survival models based on the exponential
8: 206-216. distribution. Comm. in Stat.: Theory and Methods
[51] Mahesh S. and Phoenix S. L., Lifetime 1981; 10: 1415-1427.
distributions for unidirectional fibrous composites [67] Siriteanu C. and Blostein S. D., Maximal-ratio
under creep-rupture loading. International Journal eigen-combining: a performance analysis.
of Fracture 2004; 127: 303-360. Canadian Journal of Electrical and Computer
[52] Tierney L., Asymptotic bounds on the time to Engineering 2004; 29: 15-22.
fatigue failure of bundles of fibers under local load [68] Kim I. M., Exact BER Analysis of OSTBCs in
sharing. Advances in Applied Probability 1982; Spatially Correlated MIMO Channels. IEEE
14: 95-121. Trans. on Communications 2006; 54: 1365-1373.
[53] Harlow D. G., Smith R., and Taylor H. M., The [69] Shao J. and Lamberson L. R., Modeling a shared-
asymptotic distribution of certain long composite load k-out-of-n:G system. IEEE Trans. on
cables. Tech. Report 384, Dept. of Operations Reliability 1991; 40: 205-209.
Research, Cornell Univ. 1978. [70] Akhtar S., Comment on: Modeling a shared-load
[54] Harlow D. G., Smith R. L., and Taylor H. M., k-out-of-n:G system. IEEE Trans. on Reliability
Lower tail analysis of the distribution of the 1992; 50: 189.
strength of load-sharing systems. Journal of [71] Newton J., Comment on: Modeling a shared-load
Applied Probability 1983; 20: 358-367. k-out-of-n:G system. EEE Trans. on Reliability
[55] Borges W. D. S., On the limiting distribution of 1993; 42: 140.
the failure time of fibrous materials. Advances in [72] Xie M., Poh K. L., and Dai Y. S., Computing
Applied Probability 1983; 15: 331-348. System Reliability: Models and Analysis. Kluwer
[56] Loève M., Ranking limit problem. Proc. 3rd Academic/Plenum Publishers 2004.
Berkeley Symp. Math. Stat. Prob. 1956; 2: 177- [73] Chan C. K., Availability analysis of load-sharing
194. systems. Proc. Ann. Reliability and
[57] Newman W. I. and Phoenix S. L., Time-dependent Maintainability Symp. 2003: 551-555.
fiber bundles with local load sharing. Physical [74] Lin H. H., Chen K. H., and Wang R. T., A
Review E 2001, 63: 021507. multivariate exponential shared-load model.
[58] Høyland A. and Rausand M., System Reliability IEEE Trans. on Reliability 1993; 42: 165–171.
Theory: Models and Statistical Methods. Fifth
Edition, John Wiley & Sons 1994; 158-159.
 306 Suprasad Amari, K. B. Misra, and H. Pham

[75] Shaked M., Extensions of the Freund Distribution and Analysis in Civil Engineering, Editor: Ayyub,
with Applications in Reliability Theory. B. M., CRC 1997; 99-120.
Operations Research 1984; 32: 917-925. [80] Hassett T. F., Dietrich D. L., and Szidarovszky F.,
[76] Lu J. C., Weibull extensions of the Freund and Time-varying failure rates in the availability and
Marshall-Olkin bivariate exponential models. reliability analysis of repairable systems. IEEE
IEEE Trans. on Reliability 1989; 38: 615-619. Trans. on Reliability 1995; 44: 155-160.
[77] Kececioglu D. and Jiang S., Reliability of two [81] Press W. H., Teukolsky S. A., Vetterling W. T.,
load-sharing Weibullian units. SAE Technical and Flannery B. P., Numerical Recipes in C.
Papers 1986; No: 861849. Second Edition, Cambridge University Press 1992.
[78] Liu H., Makis V., and Jardine A. K. S., Reliability [82] Dugan J.B., Tutorial: Fault-tree analysis of
Assessment of Systems Operating in Variable computer-based systems. Proc. Ann. Reliability
Conditions. Proc. ISUMA-NAFIPS 1995; 5-8. and Maintainability Symp. 2003.
[79] Liu H., Makis V., and Jardine A. K. S., [83] Amari S.V., Dill G., and Howald E., A new
Computation of Reliability for Systems Operating approach to solve dynamic fault trees. Proc. Ann.
in Varying Conditions. In Uncertainty Modeling Reliability and Maintainability Symp. 2003; 374 -
379.

View publication stats