Abstract
In this work we are focused on multi-state systems modeled by means of a special type of semi-Markov processes. The sojourn times are seen to be independent not necessarily identically distributed random variables and assumed to belong to a general class of distributions closed under extrema that includes, in addition to some discrete distributions, several typical reliability distributions like the exponential, Weibull, and Pareto. A special parametrization is proposed for the parameters describing the system, taking thus into account various types of dependencies of the parameters on the the states of the system. We obtain maximum likelihood estimators of the parameters and plug-in type estimators are furnished for the basic quantities describing the semi-Markov system under study.
Access provided by CONRICYT-eBooks. Download chapter PDF
Similar content being viewed by others
Keywords
- Multi-state system
- Reliability theory
- Survival analysis
- Semi-Markov processes
- Parameter estimation
- Time-varying model
- Scale parameter
1 Introduction
Technical and technological systems assuming multiple possible states are known as multi-state systems (MSS). Any system that is allowed to assume a finite number of performance rates can be modeled by means of a multi-state system.
Such modeling approaches, which are more realistic and provide more accurate representations of engineering systems, are much more complex and present major difficulties in system definition and performance evaluation. MSS reliability has received a substantial amount of attention in the past four decades with basic concepts being introduced in the 70s by [1,2,3,4]. Extensions and generalizations can be found in [5,6,7]. Essential achievements that were attained up to the mid 1980s are reflected in [8, 9], where one can find the state of the art in the field of MSS reliability. For theoretical advances and significant applications in MSS reliability theory in recent years, the reader is referred to [10,11,12]. For general references on continuous-time semi Markov systems and associated reliability topics, one can see [13,14,15,16,17].
Consider a process defined on a probability space \((\varOmega , \mathscr {F}, \mathbb P)\) with state space \(E=\{1,2,\ldots ,N\}\). For example, state “N” is associated with nominal performance of the system and state “1” is associated with total failure. Markov processes represent typical tools for modelling such a system. In this work we focus on multi state systems that we model by means of semi-Markov processes, which generalize typical Markov jump processes by allowing general distributions for sojourn times [13]. For this reason, the semi-Markov processes are more adapted for reliability studies (and for applications in general).
The sojourn times in a given state are assumed to belong to a general class of distributions, cf. Relation (2). The interest of this distribution class is twofold. First, it is worth noticing that the class is closed under extrema (cf. [18]) and secondly it unifies under a single umbrella, not only some discrete distributions but also and more importantly, several typical reliability distributions like the exponential, Weibull, Rayleigh and Pareto distributions.
In this chapter we consider a special case of the semi-Markov system introduced in [19]. In that article, the system under study depended on some parameters \(a_{ij},\) with i, j belonging to the state space E. In the present work, the dependence of the parameters \(a_{ij}\) on the states i and j is made explicit through a function g(i, j). We also provide several examples of such a function g that could be of interest in different modeling situations, according to the application under study. It is important to indicate that time- and state-varying parameters become quite popular since more and more systems are subject to dynamic changes. Note that the problem of time- and state-varying parameters has received in recent years, increased attention (see e.g. [20,21,22,23]) because of an ever-growing body of evidence that typical assumptions of stable parameters often appear invalid.
The chapter is organized as follows. Some preliminaries regarding semi-Markov processes are presented in Sect. 2; here we also describe the special case of the semi-Markov system introduced in [19], developed in a multi-state system framework. In Sect. 3 we describe the class of distributions considered in this work and we propose a special parametrization of these distributions. In Sect. 4 the likelihood function and the associated maximum likelihood estimators of the parameters of interest are provided.
2 A Special Case of Semi-Markov Multi-state Systems
As previously mentioned, we assume that the random system has finite state space \(E=\{1,\ldots , N\}\), \(N < \infty \) and its time evolution is governed by a stochastic process \(Z=(Z_t)_{t\in \mathbb {R}_+}.\) Let us denote by \(S=(S_n)_{n\in \mathbb N}\) the successive time points when state changes in \((Z_t)_{t\in \mathbb {R}_+} \) occur and by \(J=(J_n)_{n\in \mathbb N}\) the successive visited states at these time points. Set also \(X=(X_n)_{n\in \mathbb N}\) for the successive sojourn times in the visited states. Thus, \(X_n=S_n-S_{n-1},\) \(n \in \mathbb N^*,\) and, by convention, we set \(X_0=S_0=0.\)
Let us first recall the definition of a Markov renewal and semi-Markov process (cf. [13]). If \(( J, S) = (J_n, S_n)_{n \in \mathbb N} \) satisfies the relation
then
-
(J, S) is called a Markov renewal process (MRP);
-
\(Z=(Z_t)_{t\in \mathbb {R}_+} \) is called a semi-Markov process (SMP) associated to (J, S), where
$$ Z_t :=J_{N(t)} \quad \Leftrightarrow \quad J_{n}=Z_{S_n}, $$with
$$\begin{aligned} N(t):= \max \{n \in \mathbb {N} \mid S_n \le t \}, \; t \in \mathbb {R}_+, \end{aligned}$$(1)the counting process of the number of jumps in the time interval (0, t]. Thus, \(Z_t\) gives the state of the system at time t.
If \((J_n, S_n)_{n \in \mathbb N} \) is a MRP, it can be immediately checked that \((J_n)_{n \in \mathbb N}\) is a Markov chain, called the embedded Markov chain.
All along this work we assume that the SMP (or equivalently, the MRP) is regular, irreducible and positive-recurrent (see, e.g., [13, 24, 25] for more details on these notions).
A SM model is characterized by its initial distribution \(\alpha =(\alpha _1, \ldots , \alpha _N)\)
and by the semi-Markov kernel
Let us also introduce the transition probabilities of the embedded Markov chain \((J_n)_{n \in \mathbb N},\)
and the conditional sojourn time distribution functions
Observe that
In the sequel, we will consider a special case of semi-Markov system introduced in [19]. As it will be seen in the next section, we will consider here a particular parametrization of this system.
Let us assume that we have at our disposal a collection of positive random variables \(T_{ij},\) that can be seen as potential times spent in state i before moving (directly) to state j. We denote by \(F_{ij}(t; \theta _{ij})\) its cumulative distribution function (cdf), where \(\theta _{ij}\) is the m-dimensional parameter involved in the underlying distribution. We assume that the distribution of \(T_{ij}\) is absolutely continuous with respect to the Lebesgue measure; an associated density is denoted by \(f_{ij}(t; \theta _{ij}).\)
The dynamic of the system is as follows: the next state to be visited after state i is the one for which \(T_{il}\) is the minimum, \(l \in E\). This is the way the next state to be visited, say j, is “chosen”, namely \(j=argmin_{l \in E}(T_{il})\). Thus, for our semi-Markov system, the semi-Markov kernel becomes
where
and
which represents the cdf of the sojourn time in state i (unconditional to the next state to be visited). Note that
Let us assume that \(W_i(t)\) is absolutely continuous w.r.t. the Lebesgue measure and has a density denoted by \(f_i(t)\).
As we will be dealing in the sequel with parametric inference, whenever a quantity of interest will depend on a parameter \(\theta \in \varTheta \subset \mathbb {R}^m\), we may set this parameter as an argument. For instance, if \(Q_{ij}(t)\) depends on some parameter \(\theta ,\) we could denote it by \(Q_{ij}(t; \theta )\).
Our intention is to provide estimators of \(p_{ij}\), \(W_i(t)\), and \(Q_{ij}(t)\) under a general class of distributions, with a specific parametrization. This class of distributions and the corresponding parametrization are presented and discussed in the next section.
3 Parametric Specification of the System
The type of distributions considered for the random variables \(T_{ij}\) are first presented in this section. Then, a specific parametrization is considered. More specifically, we consider the case where the distributions \(F_{ij}(\cdot ; \theta _{ij})\), \(i, j = 1, \ldots , N,\) are of the same functional form but with different parameters, i.e., we are focusing on independent but not necessarily identically distributed (inid) random variables. Nonidentically but independently distributed random variables are usually not easy to deal with. But, when these belong to families of random variables closed under maxima or minima then elegant expressions of various statistical characteristics such as order statistics are possible. A member of such a class of distribution functions with parameter a is assumed to verify the following distributional form
Let us assume that F(x; a) is absolutely continuous w.r.t. the Lebesgue measure and let us denote its density by f(x; a), namely
The following result states that the minimum order statistic from an inid random sample from the above class has a distribution belonging to the same class.
Lemma 1
(cf. [18]) Let \(X_1, \ldots , X_N\) be inid random variables such that \(X_i \sim F(x;a_i)\) which belongs to class (2). Then the distribution function \(F^{(1)}\) of the minimum order statistic \(X_{(1)}\) belongs also to (2).
It is worth noticing that examples of distribution that belong to class (2) are the geometric distribution, the Pareto distribution, the Weibull distribution and its special cases like the exponential, the Rayleigh and the Erlang truncated exponential.
Let us now assume that the random variables \(T_{ij}\) considered in the previous section belong to the class (2), with the corresponding parameters \(a_{ij},\) i.e., the corresponding cumulative distributions \(F(t; a_{ij})\) verify (2). Moreover, we assume that we have a parametrization for \(a_{ij}\) that makes explicit the dependence on the states i and j. To be more specific, let us assume that \(a_{ij}\) has the expression
where \(a_{\infty }\) and \(e_1\) are real parameters, while g(i, j) is a known function of states i and j, depending on certain parameters. Typical examples of g(i, j) can be obtained by considering
where \(c_m, k_m, l_m, m = 1, 2,\) are real parameters. Examples of such a function g that could be of interest in different modeling situations, according to the application under study, could be:
Remark 1
-
1.
Note that this parametrization is done by analogy with a framework considered in [20], where the times between two successive failures are assumed to be inid random variables distributed according to a cumulative distribution \(F(x;a_{i})\) belonging to the class (2) with different scale parameters \(a_{i}.\) These parameters are assumed to be time varying; one type of variation along time proposed in that article is of the type \(a_{i}=a_{\infty }\left( 1-e^{-t_i/e_1} \right) , i= 1, 2, \ldots ,\) where \(t_1, t_2, \ldots \) are observed successive failure times. Nonetheless, note that, in the present chapter, the variation is on both states i and j, while in [20] the variation is along time.
-
2.
Note that, if we consider a semi-Markov system with only one state (\(E=\{1\}\)), we are in the framework of [20], where \(S_n, n=1, 2, \ldots ,\) are the successive failure times of a system, \(S_n < S_{n+1},\) and \(S_0:=0,\) while \(X_n := S_n - S_{n-1}, n=1, 2, \ldots ,\) are the times between two successive failures. It is clear that, in this case, there is no state variation anymore and a modeling like the one proposed in [20] would be appropriate.
Under these conditions, the following result concerning the main semi-Markov characteristics can be proved. For notational convenience, we set \(F(t):=F(t;1)\), \(f(t):=f(t;1)\) and \({{Q}_{ij}}\left( t; a_{ik};k=1,\ldots ,N \right) :=Q_{ij}(t).\)
Proposition 1
(cf. [19]) Under the setup of this section, the following results hold:
and
4 Maximum Likelihood Estimation
In this section we consider the problem of obtaining the maximum likelihood estimators of the parameters of the system (\(a_{\infty }\) and \(e_1,\) or, equivalently, \(a_{ij}\)). Then we will get the corresponding plug-in estimators of the main quantities defining the semi-Markov system.
Basically, two important statistical settings could be considered: either we start with one sample path, or with several sample paths. In both cases, it can be assumed that the sample paths are complete or that the sojourn time in the last visited state can be right censored (lost to follow-up, for instance). In the sequel we consider the most general case, that is the one of several sample paths with possible censored last sojourn time. The other cases can be obtained from the one we present, as a particular case; we will also give some details on this point.
Given L sample paths of a semi-Markov process censored at time M,
\(\left\{ j_0^{(l)}, x_1^{(l)}, j_1^{(l)}, x_2^{(l)}, \ldots , j_{N^l(M)}^{(l)}, u_M^{(l)}\right\} ,\) \(l=1,\ldots , L\), then the associated likelihood is
where we set
-
\(N_{i,0}^{(L)} := \sum \limits _{l=1}^{L}\mathbbm {1}_{\{J_0^{(l)}=i\}}\): the number of sample paths starting in state i;
-
\(N_i^{(l)}(M)\): the number of visits to state i up to time M of the lth trajectory, \(l=1, \ldots , L\);
-
\(N_{i}(L,M) :=\sum \limits _{l=1}^{L}N_{i}^{(l)}(M)\): the total number of visits to state i up to time M along the L trajectories;
-
\(N_{ij}^{(l)}(M)\): the number of transitions from state i to state j up to time M during the lth trajectory, \(l=1, \ldots , L\);
-
\(N_{ij}(L,M) :=\sum \limits _{l=1}^{L}N_{ij}^{(l)}(M)\): the total number of transitions from state i to state j up to time M along the L trajectories;
-
\(x_i^{(l,k)}\): the sojourn time in state i during the kth visit, \(k=1, \ldots , N_i^{(l)}(M)\) of the lth trajectory, \(l=1, \ldots , L\);
-
\(u_M^{(l)}:=M-S_{N^l(M)}\) is the observed censored time of the lth trajectory;
-
\(N_{i,M}(L)=\sum \limits _{l=1}^{L}\mathbbm {1}_{\{J^{(l)}_{N^l(M)}=i\}}\) is the number of visits of state i, as last visited state, over the L trajectories; note that \(\sum \limits _{i \in E}N_{i,M}(L)=L;\)
-
\(u_i^{(k)}\) is the observed censored sojourn time in state i during the kth visit, \(k=1,\ldots ,N_{i,M}(L)\).
Note that, for \(L = 1,\) the likelihood given in (14) reduces to the likelihood of 1 trajectory. Note also that, if the censoring time M in a certain trajectory l is a jump time, then for the corresponding observed censored time we have \(u_M^{(l)}=0\). Consequently, the contribution to the likelihood of the associated term will be equal to 1. For this reason, if no censoring is involved, the uncensored likelihood can be obtained as a particular case of (14).
For the class of distributions given in (2), the likelihood takes the form
where \(a_{ij}\) has been given in (4). Consequently, the log-likelihood has the expression
Using \(a_{ij} = a_{\infty }\left( 1-e^{g(i,j)/e_1} \right) ,\) taking the derivatives of \(\log ({\mathscr {L}})\) with respect to \(a_{\infty }\) and \(e_1\) we obtain the critical equations:
Equation (17) provides an explicit expression of \(a_{\infty }\) in terms of \(e_1\)
This expression replaced in Eq. (18) provides an equation in \(e_1\) that has to be solved numerically. Thus we obtain the corresponding MLEs \(\widehat{a}_\infty (L,M)\) and \(\widehat{e}_1(L,M)\) and also the corresponding plug-in estimator of \(a_{ij},\)
Consequently, using Proposition 1, we get the plug-in estimators of the main quantities that define the semi-Markov system, namely \(p_{ij}\), \(W_i(t)\) and \(Q_{ij}(t)\):
and
Note also that, once we have obtained the estimators of the basic quantities associated to a multi-state semi-Markov system, we can immediately obtain estimators of the associated reliability indicators, following the lines presented in [19].
Remark 2
A more general framework may be considered if some or all of the parameters involved in the function \(g(\cdot , \cdot )\) are assumed to be unknown. In such a case, the appropriate derivatives of the loglikelihood in (16) with respect to \(c_m, k_m, l_m, \ m=1,2,\) should be considered and the normal equations in addition to (17) and (18) should include the derivatives with respect to extra unknown parameters. In this more general setting, the system of equations has to be solved numerically for the estimators of the parameters to be obtained.
5 Concluding Remarks
In many settings the challenge is to determine if and where the parameters of the underlying model change their value. The rationales for time-varying parameter models may be several. For instance, the true coefficients themselves can often be viewed directly as the outcome of a stochastic process. Furthermore, even when the underlying parameters are stable, situations arise in which a time-varying coefficient approach will prove to be effective. More considerations could be provided on this topic. The present chapter deals with the problem in a general setting where a general class of distributions is considered with state-varying parameters. In particular in a multi-state system modeled by means of a special type of semi-Markov process, the parameters involved are assumed to be affected by the present state as well as the state to be visited and the likelihood together with the parameter estimates are provided under various dependency types of the parameters involved on the states of the system.
References
Barlow R, Wu A (1978) Coherent systems with multi-state components. Math Oper Res 3:275–281
El-Neveihi E, Proschan F, Sethuraman J (1978) Multistate coherent systems. J Appl Probab 15:675–688
Murchland J (1975) Fundamental concepts and relations for reliability analysis of multistate systems. In: Barlow R, Fussell J, Singpurwalla N (eds) Reliability and fault tree analysis: theoretical and applied aspects of system reliability. SIAM, Philadelphia, pp 581–618
Ross S (1979) Multivalued state component systems. Ann Probab 7:379–383
Block H, Savits T (1982) A decomposition of multistate monotone system. J Appl Probab 19:391–402
Hudson J, Kapur K (1982) Reliability theory for multistate systems with multistate components. Microelectron Reliab 22:1–7
Natvig B (1982) Two suggestions of how to define a multistate coherent system. Adv Appl Probab 14:434–455
El-Neveihi E, Proschan F (1984) Degradable systems: a survey of multistate system theory. Comm Stat Theory Method 13:405–432
Natvig B (1985) Multi-state coherent systems. In: Jonson N, Kotz S (eds) Encyclopedia of statistical sciences, vol 5. Wiley, New York, pp 732–735
Lisnianski A, Frenkel I, Ding Y (2010) Multi-state system reliability analysis and optimization for engineers and industrial managers. Springer, London
Lisnianski A, Levitin G (2003) Multi-state system reliability: assessment. Optimization and applications. World Scientific, Singapore
Natvig B (2011) Multistate systems reliability. Theory with applications. Wiley, New York
Limnios N, Oprisan G (2001) Semi-Markov processes and reliability. Birkhauser, Boston
Limnios N, Ouhbi B (2003) Empirical estimators of reliability and related functions for semi-Markov systems. In: Lindqvist B, Doksum K (eds) Mathematical and statistical methods in reliability 7. World Scientific, Singapore, pp 469–484
Limnios N, Ouhbi B (2006) Nonparametric estimation of some important indicators in reliability for semi-Markov processes. Stat Methodol 3:341–350
Ouhbi B, Limnios N (1996) Non-parametric estimation for semi-Markov kernels with application to reliability analysis. Appl Stoch Model Data Anal 12:209–220
Ouhbi B, Limnios N (1999) Non-parametric estimation for semi-Markov processes based on its hazard rate functions. Stat Infer Stoch Process 2(2):151–173
Balasubramanian K, Beg Bapat R (1991) On families of distributions closed under extrema. Sankhya Indian J Stat 53:375–388
Barbu V, Karagrigoriou A, Makrides A (2016) Semi-Markov modelling for multi-state systems. Method Comput Appl Probab
Barbu V, Karagrigoriou A, Makrides A (2017) Statistical inference for a general class of distributions with time varying parameters, submitted
Jiang R (2016) Weibull model with time-varying scale parameter for modeling failure processes of repairable systems, In: Frenkel I, Lisnianski A (eds) Proceedings of the second international symposium on stochastic models in reliability engineering. Life science and operations management (SMRLO16). Beer Sheva, Israel, pp 292–295, IEEE CPS, 15–18 Feb 2016, 978-1-4673-9941-8/16 (2016). doi:10.1109/SMRLO.2016.54
Yang T, (2011) Measurement of yield distribution: a time-varying distribution model. Agricultural and Applied Economics Association, Annual Meeting, 24–26 July 2011. Pittsburgh, Pennsylvania
Zhu Y, Goodwin B, Ghosh S (2010) Modeling yield risk under technological change: dynamic yield distributions and the U.S. crop insurance program. Dissertation Paper, Department of Agricultural and Resource Economics, North Carolina State University
Limnios N, Ouhbi B, Sadek A (2005) Empirical estimator of stationary distribution for semi-Markov processes. Comm Stat Theory Method 34(4):987–995
Moore H, Pyke R (1968) Estimation of the transition distribution of a Markov renewal process. Ann Inst Stat Math 20:411–424
Acknowledgements
The research work of Vlad Stefan Barbu was partially supported by the projects XTerM—Complex Systems, Territorial Intelligence and Mobility (2014–2018) and MOUSTIC—Random Models and Statistical, Informatics and Combinatorial Tools (2016–2019) within the Large Scale Research Networks from the Region of Normandy, France. The research work of Alex Karagrigoriou was completed as part of the research activities of the Laboratory of Statistics and Data Analysis of the Department of Mathematics of the University of the Aegean.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2018 Springer International Publishing AG
About this chapter
Cite this chapter
Barbu, V.S., Karagrigoriou, A. (2018). Modeling and Inference for Multi-state Systems. In: Lisnianski, A., Frenkel, I., Karagrigoriou, A. (eds) Recent Advances in Multi-state Systems Reliability. Springer Series in Reliability Engineering. Springer, Cham. https://doi.org/10.1007/978-3-319-63423-4_4
Download citation
DOI: https://doi.org/10.1007/978-3-319-63423-4_4
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-63422-7
Online ISBN: 978-3-319-63423-4
eBook Packages: EngineeringEngineering (R0)