Abstract
This study aims to investigate the reliability properties of the systems constituted of three identical components that are dependent of each other. Besides this, it is studied to compare between the lifetimes of such systems in the sense of hazard rate ordering. Illustrative examples are given by considering some trivariate exponential distributions that are commonly used in reliability literature for the joint distribution of the component lifetimes.
Access provided by Autonomous University of Puebla. Download conference paper PDF
Similar content being viewed by others
Keywords
- Hazard rate ordering
- Systems with dependent components
- Hazard rate function
- Exponential lifetimes
- Three-component systems
10.1 Introduction
In a reliability theory, when constructing a system model, a common assumption is to consider a system consisting of the independent components. But in many practical systems, the component’s functioning affects the others. Therefore, it is necessary to consider the lifetimes of the components to be dependent on each other. For example, a student must take a three-step test to prove his/her competency. His effort in the first stage may have been influenced by environmental conditions and stress, which will probably effect his success in the other stages. Though the independence assumption is appropriate in certain systems, many common reliability modeling practices are inappropriate. If the component lifetime is influenced by environmental exposure, then the components within a system may be dependent on each other, or this dependence may be a result of functional dependence. The study of the system reliability without independent component lifetimes is one of the more difficult problems in the reliability engineering. Though the assumption of independence can often be used to obtain joint distribution of the component lifetimes, sometimes such an assumption is questionable. So, the joint attitude of such component lifetimes needs to be modelled by convenient multivariate distribution to include this dependence structure. Multivariate exponential distributions with exponential margins are useful in reliability modeling. Because these derivations are based on shock models, and the residual lifetime of the each component is independent of its age.
In this paper, we compare the hazard rate functions of systems. Each of them consists of three dependent components whose lifetimes have trivariate exponential distributions. Furthermore, the paper gives a reliability relationship among five systems when the component lifetimes are exchangeable. An ordering relation such as hazard rate ordering between the system lifetimes assuming exponential marginals has been studied by many authors. One may refer to those works in literature such as Navarro and Shaked (2006), Navarro et al. (2006a, b), Navarro and Lai (2007), Joo and Mi (2010).
This paper is organized as follows. In Sect. 10.2, basic definitions and the reliability properties of the system with three dependent components are given. We discuss an ordering relation for possible five coherent systems. In Sect. 10.3, we illustrate the discussion by considering four trivariate exponential distributions. For the joint distribution of the component lifetimes, we assume Farlie-Gumbel-Morgenstern, Marshall-Olkin, Gumbel Type I, Gumbel-Hougaard and Sarhan-Balakrishnan type trivariate exponential distributions which are widely used in reliability literature.
10.2 Basic Tools and Main Result
We provide a concise introduction to some basic tools and notions that will be relevant for achieving the main result. Let \( X \) and \( Y \) be both random variables with survival functions \( \overline{F}_{X} \) and \( \overline{F}_{Y} \), and density functions \( f_{X} \) and \( f_{Y} \). The hazard rate function of \( X \) is defined by \( h_{X} (t) = f_{X} (t)/\overline{F}_{X} (t) \) for all \( t \) such that \( \overline{F}_{X} (t) > 0 \). The hazard rate order is defined as \( X \le_{hr} Y \Leftrightarrow h_{Y} (t) \le h_{X} (t) \) for all \( t \ge 0 \), or \( \overline{F}_{X} (t)/\overline{F}_{Y} (t) \) is increasing in \( t \). For more on the definition of the hazard rate order, readers may refer to Shaked and Shanthikumar (1994), and Lai and Xie (2006).
Let \( T_{i} \) be the lifetime of the \( i \)th system \( (i = 1, 2, \ldots, 5) \) and \( X_{i} \), \( (i = 1, 2, 3) \) stand for the lifetimes of the components. Then five systems are demonstrated respectively as \( \hbox{min} \left\{ {X_{1}, X_{2}, X_{3} } \right\} \equiv T_{1} \), \( \hbox{min} \left\{ {X_{1}, \hbox{max} \left\{ {X_{2}, X_{3} } \right\}} \right\} \equiv T_{2} \), \( \hbox{max} \left\{ {X_{1}, \hbox{min} \left\{ {X_{2}, X_{3} } \right\}} \right\} \equiv T_{3} \), \( \hbox{max} \{ \hbox{min} \left\{ {X_{1}, X_{2} } \right\},\hbox{min} \left\{ {X_{1}, X_{3} } \right\}, \) \( \hbox{min} \left\{ {X_{2}, X_{3} } \right\}\} \equiv T_{4} \), and \( \hbox{max} \{ X_{1}, \) \( X_{2}, \) \( X_{3} \} \equiv T_{5} \). Detailed discussions can be found in Barlow and Proschan (1975). Throughout the paper, \( S_{( \cdot )} \) denotes the reliability function of the system, and \( \overline{F} (x_{1}, x_{2}, x_{3} ) \) denotes the joint survival function of the component lifetimes. Then the survival functions of the five systems respectively are:
We assume that component lifetimes are identical then we reveal the following two new equations as follows:
and
Furthermore, following relation can be written for both identical and non-identical cases
From (10.1) and (10.2), we respectively write below equalities for the hazard rates of \( T_{2} \) and \( T_{3} \)
and
where, \( \alpha_{14} (t) = \frac{{S_{1} (t)}}{{S_{1} (t) + 2S_{4} (t)}} \) and \( \alpha_{45} (t) = \frac{{2S_{4} (t)}}{{S_{5} (t) + 2S_{4} (t)}} \).
Furthermore, let us consider the systems with two components which are series and parallel systems. Their hazard rates can be represented by a convex combination of the hazard rates of three-component system. Accordingly, \( T_{ss} \) and \( T_{ps} \) denote respective lifetimes of series and parallel systems. Then corresponding hazard rates is denoted by \( h_{ss} (t) \) and \( h_{ps} (t) \). From now on, we can write mentioned convex combinations as follows
and
where \( \alpha_{12} (t) = \frac{{S_{1} (t)}}{{S_{1} (t) + S_{2} (t)}} \) and \( \alpha_{35} (t) = \frac{{S_{3} (t)}}{{S_{3} (t) + S_{5} (t)}} \).
It can be seen from the Eq. (10.4) that \( h_{2} (t) \) is represented as a convex combination of \( h_{1} (t) \) and \( h_{4} (t) \). Here, it is expected and reasonable that \( h_{1} (t) \) is greater than \( h_{4} (t) \) for all \( t \). However, a counter-example was given for the coherent system by Navarro and Shaked (2006) (see Example 2.2). Now, suppose that \( h_{1} (t) > h_{4} (t) \) for some \( t \). In this case, \( h_{2} (t) \) lies between \( h_{1} (t) \) and \( h_{4} (t) \). Conversely for some \( t \), \( h_{1} (t) < h_{4} (t) \) then \( h_{2} (t) \) still lies between them. That is, if there exist an ordering amongst \( T_{1} \), \( T_{2} \) and \( T_{3} \) then these lifetimes must be ordered as \( T_{1} \le_{hr} T_{2} \le_{hr} T_{4} \) or \( T_{4} \le_{hr} T_{2} \le_{hr} T_{1} \). Similar interpretation can be made for \( T_{3} \) from the Eq. (10.5). On the other hand, according to the Eqs. (10.6) and (10.7), if one can reveal two usual orderings between \( h_{ss} (t) \) and \( h_{1} (t) \) and between \( h_{ps} (t) \) and \( h_{5} (t) \) then five lifetimes can be ordered respectively. It is easily seen that
when each five system consist of independent components. We find out the conditions when the components are dependent the five systems reveal the same ordering above. Navarro and Shaked (2006) investigated some conditions of the hazard rate order of the order statistics. Their ideas give us to consider ordering amongst for all systems with three components. To show the ordering relation amongst the five lifetimes we propose the following Lemma.
Lemma 1
\( T_{1} \le_{hr} T_{2} \le_{hr} T_{4} \le_{hr} T_{3} \le_{hr} T_{5} \) holds if one of the conditions below is satisfied:
-
(i)
both ratios \( \frac{{S_{1} (t)}}{{S_{4} (t)}} \) and \( \frac{{S_{4} (t)}}{{S_{5} (t)}} \) are decreasing function of \( t \),
or
-
(ii)
for all \( t > 0 \), \( h_{ss} (t) \le h_{1} (t) \) and \( h_{ps} (t) \ge h_{5} (t) \).
Proof
-
(i)
Suppose that the ratio \( \frac{{S_{1} (t)}}{{S_{4} (t)}} \) decreases in \( t \). Then from (ii), two facts \( \frac{{S_{2} (t)}}{{S_{4} (t)}} \downarrow t \) and \( \frac{{S_{2} (t)}}{{S_{1} (t)}} \uparrow t \) are obtained. According to the definition of the hazard rate ordering, we have \( T_{2} \le_{hr} T_{4} \) and \( T_{1} \le_{hr} T_{2} \). According to the transitivity property of the hazard rate ordering, then we get
$$ T_{1} \le_{hr} T_{2} \le_{hr} T_{4}. $$(10.9)Secondly, assume that the ratio \( \frac{{S_{4} (t)}}{{S_{5} (t)}} \) decreases in \( t \). Then from (10.3), two facts \( \frac{{S_{3} (t)}}{{S_{5} (t)}} \downarrow t \) and \( \frac{{S_{3} (t)}}{{S_{4} (t)}} \uparrow t \) are obtained. Hence, we write \( T_{3} \le_{hr} T_{5} \) and \( T_{4} \le_{hr} T_{3} \) and both relations imply
$$ T_{4} \le_{hr} T_{3} \le_{hr} T_{5} . $$(10.10)By combining (10.9) with latter relation, we get the desired result.
-
(ii)
From the Eq. (10.6), we have
$$ h_{ss} (t) - h_{1} (t) = \left( {1 - \alpha_{12} (t)} \right)\left( {h_{2} (t) - h_{1} (t)} \right) $$(10.11)This result is combined with Eq. (10.4), above equality is enhanced as
$$ h_{ss} (t) - h_{1} (t) = \left( {1 - \alpha_{12} (t)} \right)\left( {1 - \alpha_{14} (t)} \right)\left( {h_{4} (t) - h_{1} (t)} \right). $$(10.12)By the assumption \( h_{ss} (t) \le h_{1} (t) \), according to Eqs. (10.11) and (10.12) we obtain two relations such as \( h_{2} (t) \le h_{1} (t) \), and \( h_{4} (t) \le h_{1} (t) \). From the Eq. 10.4, by considering \( h_{4} (t) \le h_{1} (t) \) then \( h_{4} (t) \le h_{2} (t) \le h_{1} (t) \). Analogously we have two other relations from the Eqs. 10.7 and 10.5 that is, we have \( h_{5} (t) \le h_{3} (t) \), and \( h_{5} (t) \le h_{4} (t) \). Finally we write \( h_{5} (t) \le h_{4} (t) \le h_{1} (t) \). By considering 5 again, we have
$$ h_{5} (t) \le h_{3} (t) \le h_{4} (t) \le h_{1} (t) $$(10.13)then we get the desired result.
We investigate the results of (i) and (ii) of the Lemma 1 by considering some trivariate exponential distributions for the joint distribution of the component lifetimes.
10.3 Illustrations
We consider FGM type trivariate exponential distribution to illustrate Lemma 1, (ii). Marshall-Olkin, Gumbel Type I, Gumbel-Hougaard and Sarhan-Balakrishnan type trivariate exponential distributions which are widely used in reliability literature are considered to illustrate Lemma 1, (i).
10.3.1 Farlie Gumbel Morgenstern Distribution with Identical Marginals
Suppose component lifetimes are identically distributed as exponential with \( \lambda \) parameter. The survival distribution is given by at time t
where \( \theta \in \left( {0,\frac{1}{2}} \right) \). For more detailed discussion on this distribution can be found in Regina and Johnson (1976), Johnson and Kotz (1977) and Kotz et al. (2000).
First we compare \( h_{1} (t) \) with \( h_{ss} (t) \). To do this, we will check the monotonicity of the ratio \( S_{1} (t)/S_{ss} (t) \). Accordingly, by letting \( 1 - \exp ( - \lambda t) = u \) then the monotonicity of this ratio is equivalent to monotonicity of \( \rho_{\theta } (u) \) which is given by
We interest the sign of the first derivative of \( \rho_{\theta } (u) \) to determine monotonic property of \( \rho_{\theta } (u) \). Consequently, after some rearrangement the first derivative can be written as follows
It is seen that first two summands are negative signed. Besides, for the last summand by using the fact that \( 4\theta u\left( {1 - u} \right) \le 1 \), we have an upper bound for \( d\rho_{\theta } (u)/du \) which is given by
Hence it can be seen that \( \rho_{\theta } (u) \) decreases in \( u \). Therefore equivalently the ratio \( S_{1} (t)/S_{ss} (t) \) decreases in \( t \). As a result, \( h_{1} (t) \ge h_{ss} (t) \). Now we will compare \( h_{5} (t) \) and \( h_{ps} (t) \) in a similar way. Monotonicity of \( S_{5} (t)/S_{ps} (t) \) is equivalent to monotonicity of the \( \varphi_{\theta } (u) \) which is given by
The first derivative of this ratio can be arranged as
where \( A(u) = u(1 - u). \) Since each sums in the numerator is nonnegative the sign of \( d\varphi_{\theta } (u)/du \) is nonnegative. Therefore \( S_{5} (t)/S_{ps} (t) \) is nondecreasing function of \( t \) that is \( h_{ps} (t) \le h_{5} (t) \). Hence assumptions of Lemma 1 (ii) are fulfilled we say that when the lifetimes have a joint distribution as a special case of FGM the possible five coherent system lifetimes preserve the ordering as well as in the independence case.
10.3.2 Marshall-Olkin Trivariate Exponential Distribution
The survival function of this family is given by
where \( \lambda \)’s are corresponding shock parameters which are positive valued. For more details we refer to Marshall and Olkin (1967) and Basu (1997). According to Lemma 1, (i) we will compare \( T_{4} \) and \( T_{5} \). To do this, we check the monotonic property of the ratio \( S_{4} (t)/S_{5} (t) \) denoted by \( r_{45} (t) \). Then,
where \( \theta = \lambda_{ij} \) and \( \delta = \lambda_{123} \). Hence, we consider the numerator of the first derivative of this ratio; denoting this statement by \( g(t) \), then we obtain
with \( g(0) = - \theta \). If \( g(t) \) remains negative along the line \( \left( {0,\infty } \right) \) then its derivative may be negative signed or its maximum value has a negative sign. Therefore, we check the sign of the derivative of \( g(t) \);
As it can be seen that \( g(t) \) decreases in \( t \). Since \( g(0) = - \theta \), \( g(t) \) can not be positive valued. This implies \( r_{45} (t) \) decreases in \( t \). According to Lemma 1, (i) second condition is satisfied. Now, we consider the ratio \( S_{4} (t)/S_{1} (t) \) denoted by \( r_{41} (t) \)
We can easily say that \( r_{41} (t) \) increases in \( t \). This implies \( r_{14} (t) \) decreases in \( t \). The first condition of Lemma 1, (i) is also satisfied. As a result of Lemma 1, we have the ordering relation for this family.
10.3.3 Gumbel Type I Trivariate Exponential Distribution
The survival function of this family is given by
where \( \lambda \)’s are corresponding shock parameters which are positive valued. Some characterizations and properties of this family can be found in Galambos and Kotz (1978) and Marshall and Olkin (1967). Then we get the ratio \( r_{45} (t) \) which is given by
where \( \theta = \lambda_{ij} \) and \( \delta = \lambda_{123} \). For the simplicity,
and
are taken. Then the ratio is rewritten as
To show the monotonicity of this ratio we check the sign of the first derivative. We consider the numerator of this derivative. Denoting this statement by \( g(t) \), then we have
By using two facts that are \( A^{\prime} (t) \le \exp (A(t)) - 1 \) and \( \exp ( - B(t)) \ge 1 - B^{\prime} (t) \) then we can write an upper bound for \( g(t) \) which is given by
This implies \( r_{45} (t) \) decreases in \( t \). Now, we consider the ratio \( r_{41} (t) \)
which increases in \( t \). Hence, \( r_{14} (t) \downarrow t \). We showed that ordering relation given in Lemma 1 is valid for this family.
10.3.4 Gumbel-Hougaard Trivariate Distribution
Distribution function of this family is given by
and the survival function is
where \( m > 1 \). We refer the reader to Georges et al. (2001) for more details. Then we write the ratio
where \( b = 3^{\frac{1}{m}} \) and \( a = 2^{\frac{1}{m}} \). To show monotonic property of this ratio, we consider the following transformation: Let \( 1 - F(t)^{b} = v \) then
Now, taking second derivative of \( \varphi (v) \) we have the following result:
Hence, \( \varphi (v) \) is a convex function on \( (0, 1) \). This implies
Now, we consider the ratio \( r_{41} (t) \) which is given by
Since \( a \le b \) we say that \( r_{14} (t) \downarrow t \). For this family, we have the similar ordering relation as discussed for above two families.
10.3.5 Sarhan-Balakrishnan Type Trivariate Distribution
This family was introduced by Sarhan and Balakrishnan (2007). Franco and Vivo (2009) have determined the joint survival function of the n-dimensional extension. The survival function of this family is given by
where \( \lambda \) and \( \theta \)‘s are positive real numbers. In order to show that the monotonic property of the ratio \( r_{45} (t) \) we take the equivalence \( u = 1 - \exp ( - t) \) which simplify the notation. Thus we get the transformation \( r_{45} (u) \) as below
Since \( u^{2\theta } /1 + u^{\theta } \) increases in \( u \) this ratio decreases in \( u \). Equivalently \( S_{4} (t)/S_{5} (t) \) decreases in \( t \). By similar notation, we consider the ratio \( r_{41} (u); \)
where \( u = 1 - \exp (- t) \). This implies \( r_{14} (t) \) decreases in \( t \). We have similar ordering relation for this family.
10.4 Conclusion
The ordering relation given as the result of Lemma 1 is also valid in case of the independent component lifetimes. Because of this, in the case of dependent component lifetimes, by considering five trivariate exponential distributions for the joint distribution of component lifetimes, we emphasized that the system consisting of the dependent components satisfies the same ordering relation. In a reliability theory, for the five distribution families which have important role for the system modeling, this ordering properties obtained in this work can be seen as a useful result.
References
Barlow R, Proschan F (1975) Statistical theory of reliability and life testing. Holt, Rinehart and Winston, New York
Basu AP (1997) Multivariate exponential distributions with constant failure rates. J Multivariate Anal 61:159–169
Franco M, Vivo JM (2009) A multivariate lifetime model based on generalized exponential distributions. In: The XIII international conference, application stochastic. Model D.A., Vilnius, Lithuania, pp 378–381
Galambos J, Kotz S (1978) Characterizations of probability distributions: a unified approach with an emphasis on exponential and related models, lecture notes in mathematics, Springer-Verlag, Berlin
Georges P et al (2001) multivariate survival modelling: a unified aapproach with copulas. Groupe de Recherche Oṕerationnelle, Cŕedit Lyonnais, Working Paper
Johnson NL, Kotz S (1977) On some generalized farlie-gumbel-morgenstern distributions-II regression, correlation and further generalizations. Commun Stat Theor Meth 6(6):485–496
Joo S, Mi J (2010) Some properties of Hazard rate functions of systems with two components. J Stat Plan Infer 140(2):444–453
Kotz S et al (2000) Continuous multivariate distributions: models and applications, Second Edition edn. John Wiley and Sons, New York, p 752
Lai CD, Xie M (2006) Stochastic ageing and dependence for reliability. Springer, New York
Marshall AW, Olkin I (1967) A multivariate exponential distribution. J Am Stat Assoc 62(317):30–44
Navarro J, Shaked M (2006) Hazard rate ordering of order statistics and systems. J Appl Probab 43(2):391–408
Navarro J et al (2006a) Reliability properties of systems with exchangeable components and exponential conditional distributions. Test 15(2):471–484
Navarro J et al (2006b) Systems with exchangeable components and gumbel exponential distributions. In: Advances in distribution theory, order statistics and inference. Statistics for Industry and Technology. Birkhäuser, Boston, MA, pp 291–306
Navarro J, Lai CD (2007) Ordering properties of systems with two dependent components. Commun Stat Theor 36(3):645–655
Regina C, Elandt-Johnson (1976) Conditional failure time distributions under competing risk theory with dependent failure times and proportional hazard rates. Scand Actuar J. pp 37–51
Sarhan AM, Balakrishnan N (2007) A new class of bivariate distributions and its mixture. J Multivariate Anal 98:1508–1527
Shaked M, Shanthikumar JG (1994) Stochastic orders and their applications. Academic Press, San Diego
Acknowledgments
This paper is dedicated to the memory of our advisor emeritus Professor Yalcin Tuncer who passed away suddenly on August 18, 2011.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2015 Springer International Publishing Switzerland
About this paper
Cite this paper
Yılmaz, M., Potas, N., Topçu, B. (2015). Reliability Properties of Systems with Three Dependent Components. In: Erçetin, Ş., Banerjee, S. (eds) Chaos, Complexity and Leadership 2013. Springer Proceedings in Complexity. Springer, Cham. https://doi.org/10.1007/978-3-319-09710-7_10
Download citation
DOI: https://doi.org/10.1007/978-3-319-09710-7_10
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-09709-1
Online ISBN: 978-3-319-09710-7
eBook Packages: Physics and AstronomyPhysics and Astronomy (R0)