Probabilistic Engineering Mechanics 45 (2016) 164–176

Contents lists available at ScienceDirect

Probabilistic Engineering Mechanics

journal homepage: www.elsevier.com/locate/probengmech

Modeling and reliability analysis of systems subject to multiple sources

of degradation based on Lévy processes
J. Riascos-Ochoa a, M. Sánchez-Silva a,n, Georgia-Ann Klutke b
a
Department of Civil and Environmental Engineering, Universidad de Los Andes, Bogotá, Colombia
b
Department of Industrial and Systems Engineering, Texas A&M University, College Station, USA

art ic l e i nf o a b s t r a c t

Article history: In this paper we present a framework to model the effect of multiple sources of degradation on dete-
Received 20 August 2015 riorating systems, and to find easy-to-evaluate expressions for important reliability quantities. By con-
Accepted 23 May 2016 sidering the system deterioration as a general increasing Lévy process (i.e., a subordinator), which is a
Available online 25 May 2016
process with independent, stationary and non-negative increments, the proposed methodology allows
Keywords: the modeling of shock-based degradation (in the form of compound Poisson processes) with different
Stochastic processes distributions of shock sizes, progressive degradation as deterministic linear drift plus a Gamma process,
Deterioration and multiple sources of degradation by the superposition of any of these models into the same math-
Lifetime ematical formalism. In addition, we obtain expressions for the reliability quantities (reliability function,
Reliability
and probability density and moments of the lifetime L), and the mean and central moments of the de-
Lévy measure
terioration process Xt. Moreover, we present efficient and accurate numerical methods to compute the
Characteristic function
reliability quantities and to simulate sample paths. Several deterioration models are compared in terms
of their reliability quantities, the simulated sample paths and the feasible moments of the deterioration
Xt. Furthermore, we propose a method to mix different Lévy models that extend the available moments
with possible applications to data fitting. The results demonstrate the generality, versatility, efficiency
and accuracy of the proposed Lévy degradation framework, which can open a new productive research
field in the area of probabilistic degradation models.
& 2016 Elsevier Ltd. All rights reserved.

1. Introduction time; and degradation modeling is concerned with the description

of the time-evolution of the deterioration. Modeling degradation
The evaluation and prediction of the performance of en- requires evaluating the uncertain nature of this process [4–6]. In
gineered systems and components over time is of particular im- the literature, degradation models are usually divided into:
portance, especially to support operational decisions. For instance, (1) shock-based; (2) progressive; and (3) combined degradation.
in electro-mechanical systems, the time-dependent assessment is Shock-based degradation models consider the cumulative ef-
needed to estimate the availability and costs of operation, in- fect of sudden (and independent) events (e.g., effect of earth-
spections and maintenance actions [1]. In addition, for the parti- quakes on infrastructures or over-currents in electronic devices)
cular case of infrastructure systems, it is essential to carry out life- by removing finite amounts of the system's condition in the form
cycle cost analysis (LCCA), to make structural reliability estima- of jumps or shocks at discrete points in time [7–10]. Shock-based
tions and for directing future investments [2,3]. Evaluating the degradation can be described by two stochastic processes: the
infrastructure system performance requires that we address two inter-arrival times {Ti }i ≥ 1 (with Ti the random variable (r.v.) of the
important aspects: the identification and modeling of the de- time between the (i − 1) th and ith shock) and shock sizes {Yi }i ≥ 1,
gradation mechanisms; and the evaluation of the system's lifetime which represent the damage of each shock to the system. If after
distribution, from which important quantities such as the mean each shock damage accumulates and no maintenance is per-
and moments and the reliability function can be obtained. formed, the total deterioration by time t, Xt, can be computed as:
Let us define the system's condition at a specific point in time, t Nt
as Vt; whereby condition we mean some mechanical properties Xt = ∑ Yi,
such as the structural capacity or the stiffness. In this sense, de- i=1 (1)
terioration is defined as the decay of the system condition over
with Nt the r.v. of the number of shocks by time t. Computing the
reliability for this model involves the evaluation of the distribution
n
Corresponding author. of the sum of n random variables (i.e., shock sizes and inter-arrival

http://dx.doi.org/10.1016/j.probengmech.2016.05.002
0266-8920/& 2016 Elsevier Ltd. All rights reserved.
 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176 165

times), which requires the evaluation of n-fold convolutions, and shock model of the type CPP, and progressive models in the form
to take the limit n → ∞. In general this is a difficult problem and of stationary gamma processes or linear deterministic models are
with few exceptions Monte Carlo simulation is the only option to subclasses of Lévy processes. Moreover, as the sum of any Lévy
obtain a solution. A more tractable model assumes Ti distributed process is also Lévy, it is possible to combine any of these pro-
exponential or Phase-type (PH), and Yi distributed exponential, cesses to model combined degradation mechanisms and to mix
normal, gamma or PH, for which simpler expressions for the several CPP's with different distributions of shock sizes with no
convolutions are available (see, Iervolino et al. [6,11] and Riascos- additional difficulty. In this sense, not only the cases reported in
Ochoa et al. [12]). The assumption of Ti distributed exponential the literature can be reproduced with this formalism (e.g., cumu-
makes the process Xt a Compound Poisson Process (CPP), which is lative shock CPP plus linear drift [1]; cumulative shock CPP with
easier to follow because of its Markovian property. shock sizes distributed Gamma plus a stationary gamma process
On the other hand, in progressive (graceful) degradation, the [6]), but also we can model any arbitrary combination of Lévy
system's condition decays continuously over time. This is the case processes like the stationary gamma process with any CPP shock
of processes like the wear-out of engineering devices, fatigue of model or linear deterministic drift with gamma process.
materials, corrosion of metals, or deterioration of pavements [5,6]. The degradation model based on Lévy processes is used to find
The simplest approach to this problem is to propose a determi- analytic expressions for the mean and n-central moments of the
nistic function for the deterioration Xt. Other models propose deterioration Xt over time; as well as easy-to-evaluate formulas for
stochastic processes for Xt, like the gamma process or the Gaussian the reliability function, the probability density of the lifetime L and
process (e.g., Brownian motion), which can reproduce the tem- its mean (mean time to failure) and moments. Although there are
poral variability inherent to the degradation [13]; moreover, easy- some previous works that have used Lévy processes in deteriora-
to-evaluate expressions for the reliability quantities are available. tion (e.g., [16–18]), this paper is the first that explicitly address the
Gamma processes are more adequate because their sample paths problem of combined degradation, provides a clear definition of
are strictly increasing, which reproduces the monotonicity of de- each mechanism based on the Lévy measure, and propose efficient
terioration over time (without maintenance). In fact, gamma and accurate numerical methods to evaluate the reliability quan-
processes have been very popular in the last years as they have tities and to simulate sample paths. In addition, this paper pre-
been applied successfully in many cases (see van Noortwijk [4] for sents a method to mix different Lévy models in order to obtain
more details). greater flexibility in reproducing deterioration moments, as an
In many deteriorating systems both types of degradation me- approach to data fitting.
chanisms are present. For instance, the effect of corrosion (pro- This paper is organized as follows: Section 2 presents the
gressive degradation) and sudden events as earthquakes (shock- problem formulation and the definition of the degrading system
based degradation) in infrastructures is of particular importance and the reliability quantities into consideration. In Section 3 we
[14]. In electronic devices sudden events like over-currents and explain the basic notions of Lévy processes: the Lévy measure, the
deterioration by continuous heating of the elements are usual. It is Lévy–Khintchine formula, and the characteristic function and ex-
important to propose models for these combined degradation me- ponent of a Lévy process. We also deduce expressions for the
chanisms. In [5,15] a mathematical framework that includes both moments of a Lévy process in function of time. Section 4 is de-
mechanisms is proposed in the form of a cumulative shock model voted to the application of Lévy processes to degradation: we
(with arbitrary distributions for inter-arrival times and shock's explain the assumptions of the model, the special class of sub-
sizes) plus a deterministic function for progressive degradation. ordinators and the mathematical formulation that we propose for
However, the mathematical expressions obtained for the reliability shock-based, progressive and combined degradation. In Section 5
quantities cannot be solved numerically in an easy way (con- we illustrate the formulation with several examples of Lévy de-
volutions and integrals are involved), and Monte Carlo simulations gradation processes, and compare them in terms of their dete-
are employed. In [1] a more tractable expression for the mean of rioration moments; and also we present the method of mixed Lévy
the lifetime L is obtained, considering a linear deterministic model processes to extend the region of feasible moments. Section 6 is
for progressive deterioration and a cumulative shock model with centered in the reliability estimation of Lévy deterioration pro-
occurrences as a Poisson process and shock sizes distributed ar- cesses, by applying some inversion formulas to obtain their
bitrarily. In [6], deterioration is a combination of a gamma process probability distribution and reliability quantities. We also propose
and a shock-based model with occurrences following a Poisson a numerical method to solve these expressions and to simulate
process and shock magnitudes distributed gamma (and ex- Lévy deterioration sample paths. These results are evaluated and
ponential as an special case). Easy-to-evaluate solutions were illustrated with several examples in Section 7. Finally, the
obtained for the probability of failure (and consequently, the re- conclusions.
liability function).
However, a need still exists for proposing and solving combined
degradation mechanisms that include other cases different from 2. Problem formulation: deteriorating system and reliability
the previously presented. It is then necessary a flexible mathe- quantities
matical framework that allows the mixture of various shock pro-
cesses (i.e., with several shock size distributions), different pro- Consider a system that is placed into operation at time t¼0 and
gressive degradation models (e.g., deterministic function and whose condition decreases with time. As mentioned in Section 1
gamma processes), and the combination of any of these models to the system condition at time t is denoted by Vt, modeled as a r.v.
account for combined degradation mechanisms. More importantly, that takes values in the set of positive real numbers. Its value V0 at
this framework should provide easy-to-evaluate expressions for t¼0 is deterministic and represents the condition when the sys-
the reliability quantities that solves the numerical issues presented tem is new. The condition decreases over time due to degradation,
in their computation (e.g., convolutions and infinite integrals), and the accumulated deterioration until time t is defined by the
bringing an alternative different from Monte Carlo simulations. random variable Xt, with X0 = 0 almost surely (a.s.). We are in-
In this paper we propose a novel framework to model the effect terested in evaluating the system reliability of a system that is
of multiple sources of degradation by supposing deterioration as an abandoned after first failure and under the assumption that there
increasing Lévy process (subordinator), which is a process with is no maintenance. Hence, the condition Vt is related with the
stationary, non-negative, and independent increments. In fact, any deterioration Xt by [5]:
166 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176

Vt = max {V0 − Xt , 0} . (2) n-central moments (with n = 2, 3, 4 ) of a Lévy process Xt as a

function of time t. For more details see [20–22].
Note that the sets {Vt }t ≥ 0 and {Xt }t ≥ 0 , formed by the r.v.'s of con-
dition and deterioration for all times t ≥ 0, constitute stochastic
processes with continuous state space and continuous time. For 3.1. Definition of a Lévy process
simplicity, in this paper we will refer Xt (or Vt) as the r.v. of dete-
rioration (condition) at time t, or the stochastic process {Xt }t ≥ 0 Lévy processes are continuous-time stochastic processes with
( {Vt }t ≥ 0 ). The particular meaning will be determined by the independent and stationary increments and with right-continuous
context. sample paths having left limits. Formally, a Lévy process is defined
System failure occurs when the condition falls bellow a pre- as follows [22]: given a filtered probability space (Ω , - , , P ), an
defined threshold kn (with 0 ≤ k⁎ ≤ V0 ) that represents a service- adapted process {Xt }t ≥ 0 with X0 = 0 a.s. is a Lévy process if:
ability or operation limit for the system (i.e. a safety threshold or
limit state). The system lifetime L is a r.v. given by: 1. {Xt }t ≥ 0 has increments independent of the past; that is, Xt − Xs is
⁎ independent of -s , 0 ≤ s < t < ∞;
L = inf {t ≥ 0: Vt ≤ k ⁎} = inf {t ≥ 0: Xt ≥ V0 − k } . (3) 2. {Xt }t ≥ 0 has stationary increments; that is, Xt − Xs has the same
Two important quantities are usually required in practice: the distribution as Xt − s , 0 ≤ s < t < ∞; and
mean time to failure MTTF; and the reliability function. From the 3. Xt is continuous in probability; that is, limt → s P (Xt ∈ ·) = P (Xs ∈ ·).
proposed formulation, the MTTF is obtained by evaluating the
mean of the lifetime: 3.2. Characteristic function and characteristic exponent
MTTF =  [L ], (4)
Most of the results for Lévy processes and for the framework
where  [·] is the expectation operator. The reliability is defined as proposed in this paper are expressed in terms of the characteristic
the probability that the system accomplishes its intended function function and the characteristic exponent of the Lévy process Xt.
in a designated period of time tm (system's mission time), i.e., it is The characteristic function ϕ Xt (ω) of the r.v. Xt is defined by the
the probability that the condition Vtm at tm is greater than the limit following transformation [23], in terms of the probability density
threshold kn: P ( Xt ∈ dx ):
⁎
R (tm ) = P ( Vtm ≥ k ) = P ( Xtm ≤ V0 − k⁎). (5) ⎡ ⎤
ϕ Xt (ω)≔ ⎢ ei〈ω, Xt 〉⎥ = ∫ ei〈ω, x〉P (Xt ∈ dx),
⎣ ⎦ d (9)
Setting the mission time tm as variable, the previous expression
defines the reliability function R(t) for all t ≥ 0. where i = −1 is the imaginary unit, ω ∈ d is the transformed
In some cases the probability density function (pdf) f(t) of the variable, and 〈· , ·〉 is the inner product in d . Note that the char-
lifetime can be defined, which is given by: acteristic function contains all the probabilistic information of Xt.
Some useful properties related to characteristic functions are:
d d ⎛ ⁎⎞
f (t ) = − R (t ) = − P ⎜ Xt ≤ V0 − k ⎟,
dt dt ⎝ ⎠ (6) 1. The characteristic function ϕ Xt uniquely determines the prob-
if such derivative exists. From this expression, any of the n-mo- ability distribution P (Xt ∈ ·) and viceversa. This is known as the
inversion formula, which is explained in Section 6.1.
ments (n ≥ 1) of the lifetime can be obtained. The moments are
2. The characteristic function ϕ Xt is uniformly continuous in d .
given by:
3. For the characteristic function:
∞
 [Ln] = ∫0 f (t ) t ndt,
(7) |ϕ Xt (ω)| ≤ 1, for ω ∈ d and ϕ Xt (0) = 1. (10)

if such integral exists (i.e., converges). 4. If a random variable (stochastic process) Xt is given by a linear
It is interesting to note that the reliability function (5) can be combination of two independent random variables (stochastic
obtained from the probability distribution P ( Xt ∈ ·) of the dete- processes) Xt{1} and Xt{2}, i.e., Xt = α1Xt{1} + α2 Xt{2}, with α1, α2
rioration process Xt at time t, where  represents any measurable constants, its characteristic function ϕ Xt is given by:
set (in the sense of Lebesgue [19]) in . Note that P ( Xt ∈ ·) is a ϕ Xt (ω) = ϕ X {1} (α1ω) ϕ X {2} (α2 ω).
t t (11)
probability measure on . Then,
⁎ ⁎
R (t ) = P ( Xt ∈ ( − ∞, V0 − k ] ) = P ( Xt ∈ [0, V0 − k ] ) (8) For a Lévy process Xt, the characteristic function has a special
Note also that f(t) and  [Ln] are obtained from the reliability form. Note that Xn = X1 + (X2 − X1) + ⋯ + (Xn − Xn − 1) and the in-
function R(t). Then, if the probability distribution P (Xt ∈ ·) is ob- crements Xj − Xj − 1 (j¼2,…,n) are independent between them and
tained, the reliability quantities can be derived. One of the objec- equally distributed to X1. Hence, the distribution of Xn is obtained
tives of this paper is to model deterioration {Xt }t ≥ 0 as an increasing by the n-th convolution of the distribution of X1. Its characteristic
Lévy process and to deduce easy-to-evaluate expressions for the function can be expressed as [20]:
probability distribution P (Xt ∈ ·) and the reliability quantities. n
(
ϕ Xn (ω) = ϕ X1 (ω) . ) (12)

In general, Eq. (12) holds for any t ≥ 0, i.e:

3. Basic notions of Lévy processes
t
(
ϕ Xt (ω) = ϕ X1 (ω) . ) (13)
This section presents the basic theory of Lévy processes taking
values on d , with d = 1, 2, … . Important concepts explained in In fact, Xt can be divided in an infinite number of independent
this section are: the characteristic function and characteristic ex- and identical distributed increments as well as X1, which means
ponent of a Lévy process, the decomposition of a Lévy process, the that their distributions are infinite divisible. In particular, the
Lévy measure, and the Lévy–Kintchine formula. Furthermore, characteristic function ϕ X1 of the infinitely divisible distribution X1
herein we will present closed-form expressions for the mean and can be expressed as [20,21]:
 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176 167

ϕ X1 (ω) = e−Ψ (ω) , (14) Π (B) =  ⎡⎣ N1 (B) ⎤⎦. (20)

where Ψ is a function from d

to  (the set of complex numbers), Due to the stationary property of Lévy processes, the expected
and it is called the characteristic exponent of the distribution number of shocks Nt(B) until arbitrary time t is:
P (X1 ∈ ·). Then, from Eqs. (13) and (14), the characteristic function
 ⎡⎣ Nt (B) ⎤⎦ = Π (B) t. (21)
of Xt has the form:
ϕ Xt (ω) = e−tΨ (ω) , It is said that the Lévy measure is finite if the number of jumps
(15)
occurring in the time interval [0, 1] (or in any finite time interval)
The function Ψ is also known as the characteristic exponent of the is finite; and infinite if it is not. In the former case the jump process
Lévy process Xt. Xt{3} is a Compound Poisson Process (CPP).

3.3. The Lévy–Kintchine formula 3.6. Mean and central moments of a Lévy process

The expression for the characteristic exponent Ψ of a Lévy From the characteristic function ϕX (ω) of a random variable X it
is possible to obtain its n-moments (see, e.g., [19]):
process is given by the Lévy–Kintchine formula [20]:
 ⎡⎣ X n⎤⎦ = ( − i)nϕX(n) (0), (22)
Ψ (ω) = − i〈ω, γ 〉
1 ⎛ ⎞ where ϕX(n) (0)
denotes the n-th derivative of ϕX (ω) evaluated at
+ Q (ω) + ∫ ⎜ 1 − ei〈ω, x〉 + i〈ω, x〉1|x |< 1 ⎟ Π (dx), ω = 0. Therefore, it is possible to obtain expressions for the mo-
2 d ⎝ ⎠ (16)
ments of the Lévy process Xt for each t ≥ 0, by replacing Eq. (15)
where 1|x |< 1 is the indicator function which is equal to 1 if |x| < 1, into (22):
and 0 otherwise. In (16) γ is a vector in d ; Q (ω) is a quadratic form
⎡ ⎤ dn
on d , which means that it can be expressed as Q (ω) = 〈ω, Aω〉,  ⎢ Xtn ⎥ = ( − i)n n e−tΨ (ω) .
with A a symmetric nonnegative-definite d × d matrix [21]; and Π ⎣ ⎦ dω ω=0 (23)
is a measure on d − {0}, known as the Lévy measure, and satisfies:
Differentiating (23) for n ¼1 and using Ψ (0) = 0 from (18), the
∫ (1 ∧ |x|2 ) Π (dx) < ∞, mean of Xt takes the form:
d (17)
 [Xt ] = ie−tΨ (0) Ψ′ (0) t = iΨ′ (0) t , (24)
where ∧ indicates the min. Then, a Lévy process Xt is uniquely
determined by its characteristic triplet (γ , Q , Π ). where Ψ′(0) denotes the derivative of Ψ (ω) at ω = 0. In the same
An important property of the characteristic exponent Ψ of a way, the n-central moments μn (t ) =  ⎡⎣ (Xt −  (Xt ))n⎤⎦ of Xt can be
Lévy process is that: obtained from Eq. (23). The expressions for n = 2, 3, 4 are:

Ψ (0) = 0, (18) μ2 (t ) = Ψ (2) (0) t (25)

which can be deduced from Eq. (16) and the fact that
μ3 (t ) = − iΨ (3) (0) t (26)
Q (0) = 〈0, A0〉 = 0, or from property 3 in Section 3.2 and Eq. (14), i.
e., ϕ Xt (0) = e−Ψ (0) = 1, if only if Ψ (0) = 0.
μ4 (t ) = (3tΨ (2) (0)2 − Ψ (4) (0)) t (27)

3.4. Decomposition of a Lévy process It is clear from (24)–(26) that the mean of Xt, its variance μ2 (t )
and third central moment μ3 (t ) vary linearly with time. These
From (16) it can be shown that a Lévy process is a superposition results are important for modeling degradation and will be used in
of three independent processes [20,21]: Section 5 to compare different Lévy deterioration models.

Xt = Xt{1} + Xt{2} + Xt{3}. (19)

This formulation is referred to as the Lévy–Ito decomposition 4. Modeling degradation via Lévy processes: the Lévy de-
gradation framework
theorem and consists of three parts. The first term is a determi-
nistic (linear) process Xt{1} = γt . The second term is a Gaussian
In this section we present a framework for modeling de-
process Xt{2} with zero mean, covariance matrix A, and having
gradation based on Lévy processes, which will be referred to as the
positive and negative increments and continuous sample paths (e.
Lévy degradation framework from now on. First we mention the
g., a Brownian motion but in multiple dimensions). The third part
basic assumptions of the framework and discuss them in terms of
is a pure-jump process Xt{3}, determined by the Lévy measure Π ,
their applicability. Then, we develop the mathematical model for
and is equal to the jump process ΔXt of Xt. The condition (17) for
the degradation mechanisms of interest.
the Lévy measure Π is necessary for the jump process ΔXt does not
diverge.
4.1. Assumptions of the model

3.5. Interpretation and classification of the Lévy measure Π The Lévy degradation framework is supported on the following
assumptions:
In modeling degradation via Lévy processes, the Lévy measure
Π plays an important role. An useful interpretation is as follows 1. Deterioration is described by a 1-dimensional (1D) stochastic
[18,20]: denote Nt(B) as the number of shocks with sizes in a set B process {Xt }t ≥ 0 , with X0 = 0 a.s. (Section 2).
in d that occur until time t (i.e., ΔXt ∈ B ); note that Nt(B) is a 2. The deterioration process {Xt }t ≥ 0 has independent increments.
random variable. Then, the Lévy measure evaluated at the set B is 3. The deterioration process {Xt }t ≥ 0 has stationary increments.
equal to the expected value of the number of jumps until time t¼ 1 4. Without maintenance deterioration is increasing (i.e., not de-
with sizes in B: creasing) a.s.
168 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176

5. Multiple sources of degradation (i.e., different shock and/or

progressive processes) act independently.
Ψ (ω) = − ivω + ∫(0,∞) (1 − eiωx) Π (dx), ω ∈ R.
(30)

In summary, a subordinator Xt has the general form

The first three assumptions allow modeling deterioration as a
Xt = vt + Xt{3}, it is uniquely determined by (v, Π ) (with v ≥ 0 and
1D Lévy process, assumption 4 restricts the process to the class of
Π with support on [0, ∞)), and has characteristic exponent given
subordinators (Section 4.2), and assumption 5 allows the combi-
by the Lévy–Kintchine formula for subordinators, Eq. (30).
nation of different degradation mechanisms by using the proposed
In order to show the versatility of the Lévy framework, in the
Lévy framework. Some of these assumptions have been applied in
next sections we will present how it can be used to describe the
previous works.
two main degradation models: shock-based and progressive de-
For instance, van Noortwijk et al. in [24], considers the case in
gradation models. Also we present a general formulation to con-
which the only information available about the structure condition
struct models that describe the combined effect of both degrada-
is the amount of degradation per unit time over a finite or infinite
tion mechanisms.
time horizon. Then, there is no way to know how the cumulated
degradation was reached. It follows that the deterioration incre-
ments are exchangeable, in other words, it does not matter the 4.3. Shock-based Lévy deterioration: compound Poisson processes
order in which they appear. As a consequence, the expected de- (CPP)
terioration is linear in time. In fact, independence in the incre-
ments is a special case of exchangeability. Then, by assuming in- Shock-based degradation is caused by discrete events in time
dependence and stationarity (i.e., a Lévy process) we can model that removes finite amounts of the condition to the system (Sec-
the exchangeability property and the linear behavior of the ex- tion 1). In the proposed Lévy degradation framework, a sub-
pected deterioration with time, as shown by Eq. (24). Other ex- ordinator Wt can be used to model shock-based deterioration if the
amples are provided by Iervolino et al. in [6,11], where they following conditions are satisfied:
showed that the independence in damage increments is satisfied
in some structures by defining the kinematic ductility as an index 1. The process Xt{1} is zero (Eq. 19), i.e., the drift term v¼ 0.
of the structure seismic capacity. In [6] the stationarity property Therefore, Wt is only a jump-process Xt{3}.
holds by assuming earthquakes arrivals as a homogeneous Poisson 2. The Lévy measure ΠW of the process Wt has support on + and it
process and the aging (progressive) deterioration as a stationary is finite.
gamma process. Both, shock-based and progressive deterioration
processes are increasing and act independently on the structure. The second condition makes the process Wt a compound
Also, it is important to identify degradation processes that do Poisson process (CPP) with positive jumps (Section 3.5) and can be
not satisfy some of the assumptions. For instance, the stationarity written as:
property 3 is not achieved in age-dependent processes, i.e., pro- Nt
cesses for which the probability law of deterioration in a time Wt = Xt{3} = ∑ ΔXs = ∑ Yi,
interval depends not only on the length of the interval but also on s≤ t i=1 (31)
the initial time. This is the case of aftershock sequences after a
where Nt is the number of shocks until time t, which follows a
mainshock, whose arrival rate is nonlinear (i.e., non-stationary)
Poisson process with rate λ, and the sequence {Yi }i ≥ 1 corresponds
[11]. Relaxing the stationarity condition in the proposed Lévy
to the shock sizes (according to the definition in Section 1), which
framework defines a more general process known as non-homo-
are iid with distribution H (·) supported on [0, ∞). Therefore, the
geneous (NH) Lévy process. Finally, state-dependent processes do
Lévy measure is given by [18,20]:
not satisfy the independency property 2. This is the case in which
the probability law of deterioration at time t depends explicitly on ΠW (dx) = λH (dx). (32)
the system condition at time t. For this case other stochastic
Note that ΠW is finite because H (·) is a probability distribution (i.e.,
methods can be applied [25,26].
H (+) = 1), therefore, ΠW (+) = λH (+) = λ . The characteristic ex-
ponent ΨW of the subordinator Wt is given by:
4.2. Increasing Lévy processes: subordinators
⎛ ⎞
Formally, subordinators are processes that take values in ΨW (ω) = λ ∫(0,∞) ⎜⎝ 1 − eiωx⎟⎠ H (dx)
+≔[0, ∞) with increasing sample paths [20]. Therefore, the ⎛ ⎞
Gaussian (Brownian) component Xt{2} in Eq. (19) must be zero; i.e., = λ⎜
⎝
∫(0,∞) H (dx) − ∫(0,∞) eiωx H (dx) ⎟⎠ (33)
Q¼0. In addition, the Lévy measure Π has support on [0, ∞) (i.e.,
the process has no negative jumps); then: The first integral is equal to H (+) = 1 and the second integral
corresponds to the characteristic function ϕY (ω) of the shock sizes
∫(0,∞) (1 ∧ x) Π (dx) < ∞, (28) Yi (Section 3.2); then:

which is necessary for the sum of jumps ∑s ≤ t ΔXs to be finite. This ΨW (ω) = λ ( 1 − ϕY (ω) ) (34)
condition also implies that the Lévy measure density Π (dx ) is
and the characteristic function of Wt is:
o (1/|x|2) as x → 0. In addition, the term i〈ω, x〉1|x |< 1 = iωx1|x |< 1 in the
integral in Eq. (16) can be integrated and included as part of the ϕWt (ω) = e−λt ( 1 − ϕY (ω) ) . (35)
deterministic term −i〈ω, γ 〉 = − iωγ thanks to the condition (28),
which defines the drift coefficient of the subordinator (or just the The mean and n-central moments (n ¼2,3) of Wt are given by Eqs.
drift in [21]) as [20]: (24)–(26):
 [Wt ] = λ ( − i) ϕY′ (0) t = λ  [Y ] t (36)
v=γ− ∫(0,1) x Π (dx), (29)

that must satisfy v ≥ 0. Thus, the characteristic exponent Ψ (ω) in

μ n, W (t ) = λ ( − i)nϕY(n) (0) t = λ  [Y n] t . (37)
(16) can be expressed as:
 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176 169

The last equalities in both equations come from Eq. (22). These ϕ Zt (ω) = e−tΨZ (ω) = ( 1 − iω/β )
− λt
, (45)
corresponds to Wald's equation [27]. Note that the mean, the
second and the third central moments increase linearly with time. and the mean and n-central moments (n ¼2,3) are given by Eqs.
Section 5 presents the evaluation and comparison of these ex- (24)–(26):
pressions for several Yi distributions.  [Zt ] = λ t /β, μ n, Z (t ) = (n − 1)!λ t /β n . (46)

Note that these expressions are also proportional to t as in the CPP

4.4. Progressive Lévy deterioration
case.
The subordinator for modeling progressive deterioration can be
constructed as the sum of two independent processes: a linear 4.5. Multiple sources of deterioration with Lévy processes
deterministic drift (LD) Xt{1} = vt (with v > 0) and a jump process
In this section we extend the Lévy framework to account for the
Xt{3} with infinite and positive Lévy measure, i.e., ΠZ () = ΠZ (+) = ∞;
effect of several sources of degradation. First, we propose the
thus,
general formulation for the combination of any two Lévy models
Zt = vt + Xt{3}, (38) and finally we develop the specific case for combined degradation,
i.e., the combination of shock-based and progressive deterioration
with the additional condition (28). Note that the term in Eq. Xt{3} models.
(38), describes a jump process with infinite number of small jumps
in any finite time interval (Section 3.5). This property is adequate 4.5.1. General formulation
to model the ‘continuous’ removal of the system condition in To start, suppose two deterioration processes affecting the
progressive degradation, as we defined in Section 1. Therefore, the system, described by the subordinators Xt{1} and Xt{2}. The pro-
characteristic exponent ΨZ of the progressive deterioration Zt is cesses Xt{1} and Xt{2} can represent any shock-based or progressive
given by Eq. (30): deterioration models. To account for its combined effect by using
ΨZ (ω) = − ivω + Ψp (ω), the Lévy degradation framework, both processes should be in-
(39)
dependent. Although this condition is not general and might not
where, be satisfied in some cases, it is a convenient approximation to the
general problem [1,5,6]. Then, the total deterioration Xt is obtained
Ψp (ω)≔ ∫(0,∞) (1 − eiωx) ΠZ (dx) (40)
by the sum, which is also an increasing Lévy process:

Xt = Xt{1} + Xt{2}. (47)

is the characteristic exponent of the jump process Xt{3}. This in-
tegral cannot be split into two terms (as in Eq. (34) for the CPP) By property 4 in Section 3.2 and the expression (15) the char-
due to the infinite value of both terms (in particular, note that acteristic function of the total deterioration is given by the product
∫ ΠZ (dx ) = ∞). Then, the mean an n-central moments (n ¼2,3) of the characteristic functions of the constitutive processes Xt{1}
(0, ∞)
have the general form: and Xt{2}. Namely,

 [Zt ] = v t + iΨp′ (0) t μ n, Z (t ) = − Ψ p(n) (0)( − i)nt . (41) ϕ Xt (ω) = ϕ X {1} (ω) ϕ X {2} (ω)
t t (48)

which clearly depend on the specific Lévy measure. An example of

a Lévy process with infinite measure that has been applied for ϕ Xt (ω) = e−tΨ1(ω) e−tΨ 2 (ω) = e−t (Ψ1(ω)+ Ψ 2 (ω)) = e−tΨ (ω) , (49)
progressive degradation is the stationary gamma process GP.
where Ψ (ω), the characteristic exponent of the total deterioration
4.4.1. The gamma process Xt, is given by the sum of the characteristic exponents, Ψ1 (ω) and
The usual definition of a non-stationary gamma process Zt with Ψ2 (ω), of the processes Xt{1} and Xt{2}:
shape function Λ (t ) > 0 and scale parameter β > 0 is (see van
Ψ (ω) = Ψ1 (ω) + Ψ2 (ω). (50)
Noortwijk [4]): (1) Z0 = 0 with probability one; (2)
Zτ − Zt ∼ Ga (Λ (τ ) − Λ (t ) , β ) for all τ > t ≥ 0, with Ga the gamma Furthermore, from (24)–(26) the mean and n-central moments
distribution; and (3) Zt has independent increments. Conditions (n ¼2,3) of Xt are given by the sum of the corresponding means
(1) and (2) provide the probability distribution of a gamma pro- and n-central moments of Xt{1} and Xt{2}:
cess:
 [Xt ] =  [Xt{1} ] +  [Xt{2} ] μ n (t ) = μ n, X{1} (t ) + μ n, X{2} (t ). (51)
β Λ (t ) Λ (t )− 1
P (Zt ∈ dx) = x exp ( − βx) dx, x ≥ 0. Finally, it is easy to show that if Xt{1} and Xt{2} are constituted by
Γ (Λ (t )) (42)
jump processes with Lévy measures Π1 and Π2, respectively, the
If the shape function is Λ (t ) = λt , with λ > 0, the gamma pro- Lévy measure Π of Xt is given by their sum:
cess is a subordinator. Thus, a stationary gamma process (GP) with
Π (dx) = Π1 (dx) + Π2 (dx). (52)
rate λ and scale β, denoted here as GP (λ , β ), is a jump process with
Lévy measure density [20]:

ΠZ (dx) = λx−1e−βxdx. (43) 4.5.2. Combined degradation

Now, let us develop the framework for combined degradation.
N
Note that ΠZ is an infinite positive measure, and satisfies the re- Suppose a shock-based process Wt = ∑i =t 1 Yi with shock arrival rate
quirement (28) for a subordinator. The characteristic exponent λ (Section 4.3) and a progressive process Zt = vt + Xt{3} (Section
ΨZ (ω) and function ϕ Zt (ω) are given, respectively, by evaluating Eqs. 4.4). Recall that Yi is the ith shock size with distribution H (·)
(40) and (15): ( i = 1, 2, …) and characteristic function ϕY (ω); Xt{3} is a jump
process representing the progressive random deterioration with
ΨZ (ω) = Ψp (ω) = λ ln ( 1 − iω/β ), (44) infinite Lévy measure ΠZ and characteristic exponent Ψp (ω); and v
is the linear deterministic rate. Hence, the combined deterioration
170 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176

process Kt is given by: models in the three degradation mechanisms: shock-based or CPP
(Table 1), progressive and combined (Table 2). In addition, a
Nt ⎛ ⎞
Kt = Wt + Zt = ∑ Yi + ⎜⎜ vt + Xt{3} ⎟⎟, comparison between the different models is performed in terms of
i=1 ⎝ ⎠ (53) their moments ( μ2 (t ) and μ3 (t )) by setting a common mean de-
terioration. Finally, we propose a method to mix these basic Lévy
Let ΨW (ω) and ΨZ (ω) be the characteristic exponents of the models in order to increase the region of feasible moments. The
processes Wt and Zt, respectively, which are given by Eqs. (34) and results can help to further research into fitting Lévy models to
(39). Assuming independency between these two processes and deterioration data.
by (50) the characteristic exponent ΨK (ω) of the combined process
Kt is given by: 5.1. Expressions for specific models
ΨK (ω) = ΨW (ω) + ΨZ (ω) = λ (1 − ϕY (ω)) + ( − ivω + Ψp (ω)).
Recall from (34) and (35) that the characteristic exponent
The characteristic function ϕ Kt (ω) is given by the product of the ΨW (ω) and function ϕWt (ω) for a shock-based (CPP) model Wt re-
characteristic functions: quire the expression for the characteristic function ϕY (ω) of the iid
shock sizes Yi (i = 1, 2, …). Thus, we study the following dis-
ϕ Kt (ω) = e−λt (1 − ϕY (ω)) eivtω − tΨp (ω) , (54) tributions for Yi which posses closed-form expressions for ϕY (ω):
and the mean and n-central moments (n¼ 2,3) of Kt are given by Dirac delta δ (y ) (with mean y), Exponential Exp (β ) (rate β), Uni-
the sum of the corresponding means and n-central moments for form U (y − a, y + a) (with mean y, and a < y ), Log-Normal
each mechanism: LN (μ, σ ), and Phase-type (PH) distribution with representation
PH (τ , T). Recently in [12], Riascos-Ochoa et al. have shown that PH
 [Kt ] =  [Wt ] +  [Zt ] = λ  [Y ] t + vt + iΨp′ (0) t (55) distributions offer a general and versatile alternative to model
shock-based degradation because of their matrix-analytic property
and denseness in the set of positive distributions [28]. In the
μ n (t ) = μ n, W (t ) + μ n, Z (t ) = λ  [Y n] t − Ψ p(n) (0)( − i)n t , (56) proposed Lévy framework this distribution is easily accom-
modated because it possesses a closed-form expression for its
where we have used (36), (37) and (41). characteristic function and moments. Table 1 also shows formulas
Finally, by (52) the Lévy measure ΠK associated to the com- for the mean  [Y ], coefficient of variation
bined process Kt is given by:
cov (Y ) =  [Y2] −  [Y ]2 / [Y ], second moment  [Y2] and third
ΠK (dx) = ΠW (dx) + ΠZ (dx) = λH (dx) + ΠZ (dx). (57) moment  [Y3] of the Yi distributions considered, which are ne-
cessary in (36) and (37) to obtain the mean and central moments
where ΠW and ΠZ are the Lévy measures of the shock-based and of Wt. In all the models the shock arrival (Poisson) rate is λ.
progressive processes. Besides, the cases studied in Table 2 for progressive degrada-
tion are: Linear deterministic (LD) with drift v, and the stationary
gamma process (GP) with parameters (λg , βg ). The combined de-
5. Specific models and feasible moments: an approach to data gradation models are: the LD model combined with each CPP
fitting model and the GP combined with each CPP.

The objective of this section is to show explicit expressions for 5.2. Feasible moments
the characteristic exponent Ψ (λ ), mean  [Xt ] and second and third
central moments, μ2 (t ) and μ3 (t ), for several Lévy deterioration For comparison purposes, the parameters of the different Lévy

Table 1
Examples of shock-based (CPP) Lévy models along with the expressions for the characteristic function ϕY (ω) , mean  [Y ] and moments  [Y n] (n¼ 2,3) for the shock sizes Yi
and characteristic exponent Ψ W (ω) , mean  [Wt ] and n-central moments μn, W (t ) (n¼ 2,3) for the corresponding shock-based Lévy deterioration Wt.

Shock-based (CPP) models

Quantities for Yi and Wt Delta Yi ∼ δ (y ) Uniform Yi ∼ U (y − a, y + a) Exponential Yi ∼ Exp (β ) Log-normal Yi ∼ LN (μ, σ ) PH-type Yi ∼ PH (τ , T)

ϕ Y (ω ) eiωy eiω (y + a) −eiω (y − a) 1

ϕYLN (ω) a −τ (T + iωI)−1t
iω2a 1 −iω /β
 [Y ] y y y = 1/β 2/2 y = − τ T −11
y = eμ+σ
cov (Y ) 0 a/ 3 y 1 σ2 2τ T−21 − (τ T−11)2
e −1
−τ T−11
 [Y2] y2 y2 + a2/3 2y2 2τ T −21
(
y2 cov (Y )2 + 1)
3 3
 [Y3] y ya2 + y 3 6y 3 −6τ T −31
y ( cov (Y ) + 1)
2 2

Ψ W (ω ) λ (1 − ϕY (ω)) λ (1 − ϕY (ω)) λ (1 − ϕY (ω)) λ (1 − ϕY (ω)) λ (1 − ϕY (ω))

 [Wt ] λty λty λty λty λty
μ2, W (t ) λty2 λt (y2 + a2/3) 2λty2 λt (2τ T −21)
(
λty2 cov (Y )2 + 1 )
μ3, W (t ) λty 3 λt (ya2 y 3) 6λty 3 3 λt ( − 6τ T −31)
+ λty2 ( cov (Y )2 + 1)
2
(iω)n nμ + n2 σ
a
ϕYLN (ω) = ∑∞
n=0 e 2 , which is not a closed-form and convergent expression. Hence, we use the approximation [29]:
n!

μ − 1 σ 2 (ωe μ) 2 + 1 iσ 4 (ωe μ)3 − 2σ 6 (ωe μ)4 +⋯ ⎛ 3 6 8 ⎞−1/2

ϕYLN (ω) ≈ eiωe e 2 2 ⎜1 − iσ 2ωe μ + σ 4 (ωe μ) 2 + iσ (ωe μ) 3 − σ 8 (ωe μ)4 + ⋯⎟ .
⎝ 2 3 ⎠
 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176 171

Table 2
Examples of progressive and combined Lévy models along with the expressions for the characteristic exponent Ψ (ω), mean  [Xt ] and n-central moments μn (t ) (n¼ 2,3) for
the corresponding Lévy deterioration Xt.

Quantities for Xt Progressive models Combined models

LD (Drift v) Gamma process GP (λg , βg ) Shocks þ LD Shocks þ Gamma process

Ψ (ω ) −ivω λg ln (1 − iω/βg ) λ (1 − ϕY (ω)) − ivω λ (1 − ϕY (ω)) + λg ln (1 − iω/βg )

 [Xt ] vt λg t/βg λt  [Y ] + vt λt  [Y ] + λg t/βg
μ2 (t ) 0 λg t/βg2 λt  [Y2] λt  [Y2] + λg t/βg2
μ3 (t ) 0 2λg t /βg3 λt  [Y3] λt  [Y3] + λg t/βg3

deterioration models are defined in order to have the same mean model with shock sizes Yi with positive support. As the set of PH
deterioration  [Xt ]. This allows to compare different Lévy models distributions is dense in this set of positive distributions (see e.g.,
with identical deterioration trends by their second and third- [12,28]), the shadowed region is also feasible for the CPP model
central moments. For the CPP models it is defined the same mean with Yi distributed PH (denoted as CPP-PH). The degenerate cases
y =  [Y ] for all the Yi distributions considered in Table 1, then lie on the vertical line μ2 (t ) /t = λE [Y ]2 = λy2, with Yi ∼ δ (y ) (CPP-
 [Wt ] = λty for all the models. For the GP model with deterioration Delta) as the dot in the lower limit of the region. The non-
Zt, the rate λg and scale parameter βg are defined such that its degenerate case CPP with Yi distributed Exponential (CPP-Exp) is
mean deterioration  [Zt ] = λg t /βg = λty =  [Wt ], and for the LD the interior point with coordinates (2λy2 , 6λy 3 ). Note that the cases
model the drift is defined as v = λy . With these definitions the CPP with Yi distributed Uniform (CPP-Unif) and Log-Normal (CPP-
mean deterioration is equal for all the models:  [Xt ] = λty . LN) are 1-dimensional curves because both distributions have two
Since the second and third-central moments are proportional parameters, one of which fixes the mean y, and the other de-
to time, the quantities μ2 (t ) /t and μ3 (t ) /t are proportionality termines different second and third moments (the parameter a in
constants that represent the increasing rates of the moments with the Uniform distribution and s(or equivalently cov(Y)) in the Log-
time. By plotting these rates for each Lévy deterioration model as Normal). Finally, the progressive model LD is the point at the
ordered pairs (μ2 (t ) /t , μ3 (t ) /t ) in a plane, it is possible to visualize origin, and the GP model follows the quadratic curve in Fig. 1.
them and identify the feasible moments for each model (Fig. 1). By
analyzing Tables 1 and 2, it is clear that a natural dimension for the 5.3. Mixed Lévy models: an approach to data fitting
rate μ2 (t ) /t is λy2 in all the models, and for μ3 (t ) /t it is λy 3. These
are the dimensions used for plotting the rates in Fig. 1. In some cases there is no knowledge of the degradation me-
Recall from Eq. (37) that for the CPP models μ2, W (t ) /t = λ  [Y2] chanisms affecting a system (call it artifact), due to lack of in-
and μ3, W (t ) /t = λ  [Y3], then, the accessible region for the CPP formation about the physics behind the process, the external
models in Fig. 1 depends on the particular Yi-distribution and the conditions or the system itself; or because the complexity of some
feasible values of  [Y2] and  [Y3]. It can be demonstrated [30] that or all of these factors. Therefore, it is not possible an accurate
these values satisfy  [Y3] ≥  [Y2]2 / [Y ] and  [Y2] > E [Y ]2 for identification of the degradation mechanisms involved (shock-
nondegenerate distributions with support on [0, ∞) and finite based, progressive or combined) or their properties (e.g., shock
moments. This defines the shadowed region at right in Fig. 1 size distributions, shock arrival rate, etc.). Instead, we can obtain
which represents the accessible region of moments for any CPP deterioration data at one or several inspection times for one or
several deteriorating artifacts, from which we can evaluate the
deterioration moments in function of time. Hence, it would be
very useful a moment-fitting method to approximate a Lévy de-
terioration model to the dataset. In this section we propose a
method that can help in this aspect which in simple terms consists
in extending the region of feasible moments of the individual Lévy
models in Fig. 1 by defining mixtures of these basic models. Recall
that the applicability of this method requires that the mean and
second and third central moments of the deterioration data are
linear (or closely linear) in time, in order to approximate the actual
deterioration to a Lévy deterioration model.
Suppose we have two independent Lévy processes Xt{1} and Xt{2}
with closed-form expressions for their characteristic exponents,
Ψ1 (ω) and Ψ2 (ω), respectively. In other words, Xt{1} and Xt{2} can be
any shock-based, progressive or combined model from Table 1 or
2. A new Lévy model Xt can be proposed by defining its char-
acteristic exponent Ψ (ω) as a linear combination of the char-
acteristic exponents of the basic models Xt{1} and Xt{2}:
Ψ (ω) = α1Ψ1 (ω) + α2 Ψ2 (ω), (58)

where α1, α2 > 0 are the superposition coefficients. Let us name

Fig. 1. Feasible second and third central moments (in terms of their temporal rates
μ2 (t )/t and μ3 (t )/t ) for several Lévy deterioration models in Tables 1 and 2, and the process Xt such defined a mixed Lévy deterioration model to
mixed Lévy models (Section 5.3) with common mean deterioration  [Xt ] = λyt . The differentiate it from the combined deterioration model in Section
axis is in natural units of λy2 and λy 3, with λ the Poisson rate of the shock arrivals in 4.5.
the CPP's models, and y the mean shock size. It is clear that the mean and n-central moments (n ¼2,3) are
172 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176

given by the linear combination of the corresponding means and deterioration process Xt, and other key reliability quantities it is
n-central moments for the constitutive processes: necessary to invert Eq. (9) to obtain P (Xt ∈ ·) from the character-
istic function ϕ Xt (ω) of Xt.
 [Xt ] = α1 [Xt{1} ] + α 2  [Xt{2} ] μ n (t ) = α1μ n, X{1} (t ) + α 2 μ n, X{2} (t ). (59)
Following [19], for the random variable Xt, with a < x :
Note from Eq. (59) that we can obtain several values for the 1⎡ ⎤ 1 ∞ −iωa −iωx
mean and moments of deterioration, different from those asso- P (X t ∈ (a , x ) ) + ⎢ P (X t = a ) + P (X t = x ) ⎥ =
2 ⎢⎣ ⎦⎥
∫ e −ω e ϕ Xt (ω) dω.
2πi −∞ (61)
ciated to the constitutive processes Xt{1} and Xt{2}, by varying the
superposition coefficients α1 and α2. In fact, note that the ex- In cases like the CPP with shock sizes following discrete dis-
pressions (50) and (51) associated to the sum of two Lévy in- tributions, the distribution P (Xt ∈ ·) is discontinuous in the points
dependent processes (Section 4.5) are particular cases of (58) and (atoms) of the distribution. In the rest of values the distribution is
(59) for a mixed Lévy deterioration model, obtained by setting continuous. If a and x are not atoms, P (Xt = a) = P (Xt = x ) = 0 and
α1 = α2 = 1. The conceptual difference in a mixed model is that, P (Xt ∈ (a, x )) = P (Xt ∈ (a, x]), then:
although the constitutive processes Xt{1} and Xt{2} are also defined a
1 ∞ e−iωa − e−iωx
priori as in a combined degradation model, there is no certainty P (Xt ∈ (a, x]) =
2π i
∫−∞ ω
ϕ Xt (ω) dω.
(62)
that these processes or their mixture Xt constitute the actual de-
terioration affecting the system (as opposite to the combined de- From this expression the cumulative distribution function
gradation model). The objective then is that the mixed Lévy model P (Xt ∈ ( − ∞ , x]) can be obtained [31]:
Xt be as close as possible to the actual deterioration process. This is
1 1 ∞ e−iωx
accomplished by varying the superposition coefficients α1 and α2 P (Xt ∈ ( − ∞, x]) = − ∫−∞ ϕ Xt (ω) dω,
2 2π i ω (63)
to get a closer approach of the mean and moments of Xt, Eq. (59),
to the actual mean and moments of the deterioration data.
To illustrate the flexibility of a mixed Lévy deterioration model, 6.2. Expressions for the reliability quantities
consider the mixture Xt of any two processes Xt{1} and Xt{2} studied
in Section 5.2 whose moments were plotted in Fig. 1. Suppose that The last equation corresponds to the reliability function R(t) for
the parameters λand  [Y ] = y are adjusted so that the means x = V0 − k⁎ in Eq. (8). In this equation, by changing the notation
 [Xt{1} ] =  [Xt{2} ] = λyt correspond to the mean of the deterioration and writing R(t) as Rx(t), to indicate that x is the deterioration to be
data. Therefore, for the mixed process Xt to have the same mean surpassed for a system's failure, the reliability is given by:
 [Xt ] = λyt , the superposition coefficients must satisfy:
1 1 ∞ e−iωx
α1 + α2 = 1 (60) Rx (t ) =
2
−
2π i
∫−∞ ω
ϕ Xt (ω) dω

as it is clear from (59). Hence, the expressions (58) and (59) for the 1 1 ∞ e − i ω x − t Ψ (ω )
characteristic exponent Ψ (ω) and the second and third central
=
2
−
2π i
∫−∞ ω
e dω
(64)
moments of Xt, μ2 (t ) and μ3 (t ), are obtained by the convex mixture
Differentiating Eq. (64) with respect to t, we obtain the lifetime
of the corresponding expressions of the original processes Xt{1} and density:
Xt{2}.
In Fig. 1 some mixed processes are represented in the form of 1 ∞ e−iωx
straight dashed lines joining the corresponding points of the ori-
fx (t ) = −
2π i
∫−∞ ω
Ψ (ω) e−tΨ (ω) dω
(65)
ginal processes. We have plotted the mixture of two different CPP
models (e.g., CPP-LN and CPP-Exp), and the mixture of the pro-
6.3. Numerical solution
gressive models with some CPP models (e.g., GP with CPP-Delta,
LD with CPP-PH, GP with CPP-PH ). As shown in Fig. 1 these mixed
The expressions for the reliability quantities must be evaluated
models extend the region of feasible moments to include the
numerically by approximating the improper integrals (64) and
shadowed region at left. These results demonstrate the versatility
(65). The usual methods used for this purpose are based in the
of the proposed mixed Lévy model to fit moments of deterioration
discretization and truncation of these integrals and to approximate
data.
them by finite sums [32]. In the literature there are several rules of
discretization out of which simpler rules such as the trapezoidal
rule [33,34,32,35] have shown to be accurate and efficient. Then,
6. Reliability and lifetime distribution in the Lévy degradation
the expression for the reliability function (64) can be approxi-
model
mated by [33]:
∞
In this section we present numerical methods to compute the 1 1 e−ix (m − 1/2) h −tΨ ((m − 1/2) h)
reliability quantities (reliability function, lifetime density and its
Rx (t ) ≈ Rx (t; h)≔ − ∑ e
2 2π i m =−∞
(m − 1/2) (66)
moments) and sample paths of an arbitrary Lévy deterioration
process Xt. An important advantage of these methods is that they where h > 0 is the discretization step size. It is necessary to
overcome the analytical difficulties when evaluating the reliability truncate the infinite sum in (66) at a maximum/minimum index
in shock-based deterioration (e.g., infinite sums and convolutions) 7M; then:
or in multiple sources of deterioration, allowing the efficient and M
accurate computation of shock-based processes (with shock sizes 1 1 e−ix (m − 1/2) h −tΨ ((m − 1/2) h)
Rx (t ) ≈ Rx (t; h, M )≔ − ∑ e
having closed-form expressions for their characteristic function, e. 2 2π i m =−M
(m − 1/2) (67)
g., Table 1) and the combination of any of these CPP models with
Similar expressions are obtained for the density:
any progressive model (linear or gamma process).
M
1 e−ix (m − 1/2) h
6.1. Inversion formula fx (t ) ≈ fx (t; h, M )≔ − ∑ Ψ ((m − 1/2) h)
2π i m =−M
(m − 1/2)

In order to derive the probability distribution P (Xt ∈ ·) of a Lévy e−tΨ ((m − 1/2) h) . (68)
 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176 173

Note that the general form of the terms into the summations in example, the increment-sampling method described in [4]. Thus,
(67) and (68) is gm (x, t ) . In both expressions |gm (x, t )| is bounded by considering that Lévy processes have independent and identically
(m −1 /2)
1, hence, the terms decrease at least at a rate 1/m. Therefore, a distributed increments (iid), the procedure consists of three steps:
good approximation is to take:
1. Define a time increment Δt > 0 and the set of times t1, t2, … , tn
M ≈ 105 (69) with ti = ti − 1 + Δt and t0 = 0 at which damage increments
^ ^
Xi = (Xti − Xti − 1) will be evaluated. Note that each Xi is iid to
The discretization step size h is determined according to the
Xt1 − Xt0 = Xt1 = XΔt .
period of oscillation of the summands as a function of ω. In the
2. The cumulative distribution function (CDF) of XΔt , given by
expression for the reliability Rx(t), Eq. (64), the oscillations depend
P (XΔt < x ) = Rx (Δt ), is numerically computed with Eq. (67) for
on the product of e−ixω and the characteristic function e−tΨ (ω). For
fixed t = Δt and several values of x.
the first factor it is clear that the associated frequency of oscillation
3. An inverse transform sampling is applied to generate the n
is x/2π . For the second factor it depends on the specific char-
random increments X^i from the CDF. The sample path is con-
acteristic exponent Ψ (ω) which can have many components in
structed as a series of successive increments X^i occurring at
frequency. However, by definition:  [eiXt ω] = e−tΨ (ω), then a first
times ti.
order approximation is to take  [eiXt ω] ≈ eiω [Xt ], which suggests an
approximate frequency equal to  [Xt ] /2π . In other words, the
frequency is proportional to the mean deterioration at t. The fre-
7. Illustrative examples
quency of the product is approximated by the sum of the fre-
quencies: ω⁎ = (x +  [Xt ]) /2π . With these considerations the dis-
This section shows results from the numerical computation of
cretization step size can be taken equal as a fraction r of the as-
the reliability quantities and sample paths of several Lévy dete-
sociated period 1/ωn:
rioration models. We first present sample paths of shock-based,
2π 2π progressive (GP) and combined deterioration, and finally we show
h = h (x, t; r )≔r =r ,
x +  [Xt ] x + tΨ′ (0) i (70) a validation of the numerical computations by comparing the
calculated pdf's of the lifetime L and its moments with available
where the last equality comes from Eq. (24). The factor r, which
expressions for particular cases. In all the numerical experiments,
satisfies 0 < r < 1, can be decreased or increased in order to have a
the truncation level in Eqs. (67) and (68) was M = 105, and the
better accuracy or better efficiency, respectively. Note that the last
discretization h = h (x, t; r ) was taken as described in (71) with
expression depends on the characteristic exponent Ψ (ω) of the
r = 1/20.
process, and on the time t and deterioration level x in which the
reliability quantities are computed. To avoid large values of h for
7.1. Example 1: progressive degradation
small x and t, an additional term equal to the mean deterioration
at t¼ 1 is summed into the denominator, to give:
Let's construct several sample paths of a system subjected to
2π 2π two progressive deterioration processes, Zt{1} and Zt{2}, using the
h (x, t; r ) = r =r
x +  [Xt ] +  [X1] x + (t + 1) Ψ′ (0) i (71) Lévy framework proposed. In the first case, degradation is mod-
eled using a gamma process GP1 (λg = 1, βg = 1/2) and for the sec-
The numerical examples in Section 7 use a value r = 1/20. The
ond GP2 (λg = 0.2, βg = 1/10). Note that, according to the expres-
computations were performed in the software MATLAB®.
sions in Table 1, the mean deterioration for both models is
 [Zt{1} ] =  [Zt{2} ] = 2t .
6.4. Numerical computation of the moments of lifetime
The sample paths of both processes are shown in Fig. 2. Si-
mulation was implemented by using the method described in
The moments of the lifetime  [Ln] can be approximated by
Section 6.5 with a time increment Δt = 0.1. The mean of the de-
numerically integrating Eq. (7) by, say, the trapezoidal rule. The
terioration process is indicated with a dashed line and, as ex-
procedure consists of two steps:
pected, in both cases the sample paths are distributed around it.
Note that they have sections of ‘continuous’ increasing deteriora-
1. Define a time increment Δt > 0 and the set of times t1, t2 ,..., tN
tion (the small-jumps contribution in the Lévy measure) and
with ti = ti − 1 + Δt and t0 = 0 at which the pdf fx(t) of the lifetime
sections with visible jumps (big shocks).
L is evaluated by using the approximation fx (t; h, M ) from Eq.
(68). The final time tN and the increment Δt are set in order to
7.2. Example 2: shock-based degradation
have the trapezoidal approximation
∞
∫0 fx (t ) dt ≈ F x (Δt , t N ) In this example we are interested to draw several realizations
of two shock-based models using the Lévy framework. The models
t − t0 ⎡ ⎤
≔ n ⎢ f (t 0 ) + 2fx (t1) + 2fx (t 2 ) + ⋯ + 2fx (t N − 1) + fx (t N ) ⎥⎦ are: (1) CPP-Delta Wt{1} with shock sizes Yi ∼ δ (y = 10), and (2) CPP-
2N ⎣ x (72)
Exp Wt{2} with Yi ∼ Exp (1/β ) and rate β = 1/10. In both cases the
close to 1 with an absolute error 1 − Fx (Δt , tN ) less than a rate of shock occurrence is λ = 0.2. As in the previous
predefined error ϵ. The examples in Section 7.4 show good example both models have the same mean deterioration
results with ϵ = 10−4 .  [Wt{1} ] =  [Wt{2} ] = 2t .
2. Approximate the moments  [Ln] by applying the trapezoidal The sample paths are shown in Fig. 3 with a time increment
rule to the integral (7): Δt = 0.1. It can be observed that while the CPP-Delta model has
t − t0 ⎡ n
 [Ln] ≈ N
⎤ always shocks with identical sizes, i.e., y¼ 10, shocks in the CPP-
⎢ t f (t 0 ) + 2t1n fx (t1) + 2t2n fx (t 2 ) + ⋯ + 2tN
n f (t n
− 1 x N − 1) + tN fx (t N ) ⎥ .
(73)
2N ⎣ 0 x ⎦
Exp model have different sizes. As expected, in both cases the
sample paths are distributed around the dashed line that re-
6.5. Construction of sample paths by using simulation presents the mean. Note that the dispersion around the mean is
greater for the CPP-Exp model, which is explained from the fact
Sample paths of different Lévy deterioration processes can be that its second central moment, μ2 (t ) = 2ωty2 = 40t , is larger than
simulated from its probability distribution P (Xt ∈ ·) using, for the one for the CPP-Delta model, μ2 (t ) = ωty2 = 20t (see
174 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176

150 150

Deterioration Z{2}
{1}

t
Deterioration Zt

100 100

50 50

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Time (years) Time (years)
Fig. 2. Sample paths of progressive deterioration models described by gamma processes: (a) Process Zt{1} with parameters GP1 (λg = 1, βg = 1/2) and (b) Process Zt{2} given by
GP 2 (λg = 0.2, βg = 1/10). The trend in both models is  [Zt{1} ] =  [Zt{1} ] = 2t .

Tables 1 and 2). 7.4. Lifetime density and moments estimation

7.3. Example 3: combined degradation Consider a system that degrades with failure threshold x ¼99.
We are now interested in obtaining the pdf fx(t) and moments  [Ln]
In this last example, we are interested in the sample paths of of the lifetime L for different degradation models, by applying Eqs.
two combined deterioration processes. The first process Kt{1} is a (68) and (73).
combination of a CPP model with shock sizes distributed uni- For the case of shocks we consider several CPP models with
formly between 8 and 12 (i.e., mean  [Y ] = y = 10 and a ¼2 in rate λ = 0.1 and the following shock sizes: (1) Yi distributed
Table 1) and a linear deterministic (LD) model with drift v ¼2. The δ (y = 20); (2) Exp (β = 1/20); (3) U (0, 40); and (4) LN (μ, σ ) with
second process Kt{2} is constituted by a progressive degradation parameters (μ, σ ) such that the mean of shock sizes is  [Y ] = 20
given by the gamma process Zt{1} from Example 1 and the CPP- with a coefficient of variation COV (Y ) = 2, according to Table 1. On
Delta model Wt{1} presented in Example 2. The Poisson rate for the the other hand, progressive degradation is modeled as a GP with
CPP models is λ = 0.2. parameters GP (λg = 0.1, βg = 1/20). The mean deterioration in all
Several realizations of Wt{1} and Wt{2} are shown in Fig. 4. It is the models is  [Xt ] = 2t .
clear that the sample paths are composed by sections of pro- Fig. 5(a) sketches fx(t) and Table 3 shows the lifetime moments
gressive (graceful) deterioration and jumps due to the shock de-  [Ln] ( n = 1, 2, 3) for the previous models. The time increment is
gradation. For Wt{1} the linear behavior of the progressive con- Δt = 1, with different values of the final time tN, with their corre-
tribution is adequately reproduced. For Wt{2} the progressive part is sponding error 1 − Fx (Δt , tN ) (Section 7.4). First to note, although
more random due to the nature of the gamma process. Finally, the mean deterioration in all the models is the same, i.e.
note that for both processes the mean deterioration is  [Xt ] = 2t , the lifetime distribution and its moments vary. The
 [Wt{1} ] =  [Wt{2} ] = 4t , represented by the dashed lines around reason is that the variances of the deterioration μ2 (t ) =  [Xt2 ] for
which the sample paths are distributed. each process are different, as can be seen in Tables 1 and 2:

150 150
Deterioration Wt{1}

Deterioration W{2}
t

100 100

50 50

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Time (years) Time (years)
Fig. 3. Sample paths for: (a) CPP-Delta model Wt{1} with shock sizes Yi ∼ δ (y = 10), and (b) CPP-Exp model Wt{2} with Yi ∼ Exp (β = 1/10) . The Poisson rate is λ = 0.2.
 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176 175

300 300

{2}
{1}

Deterioration Kt
200
t

200
Deterioration K

100 100

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Time (years) Time (years)
Fig. 4. Sample paths for two combined degradation models: (a) Kt{1} as a linear deterministic model (with drift v¼ 2) plus a CPP model with Yi ∼ U (10 − 2, 10 + 2) , and (b)
Kt{2} a gamma process with parameters GP1 (λg = 1, βg = 1/2) plus a CPP model with Yi ∼ Exp (β = 1/10) . The Poisson rate is λ = 0.2.

μ2 (t ) = 40t , (160/3) t , 80t , 120t for the CPP-Delta (and GP model), Table 3
CPP-U, CPP-Exp, and CPP-LN, respectively. Note in Fig. 5(a) and Numerically computed moments of the lifetime L with the method in Section 7.4.
The percentage errors for the GP, CPP-Delta and CPP-LN take as reference the values
Table 3 that the processes follow this same order according to the
obtained from the theoretical expressions (74).
mean  [L ] or variance  [L2] of the lifetime L. In conclusion, the
processes with greater variance produce greater means and var- Lévy model GP CPP-Delta CPP-U CPP-Exp CPP-LN
iances in their lifetime.
Time increment: Δt 1.0 1.0 1.0 1.0 1.0
In addition, the results for the GP model, the CPP-Delta and the
Final time: tN 300 300 400 400 500
CPP-Exp models were compared with available expressions for the
Error ϵ = |1 − Fx (Δt , tN )|
lifetime density (see [4] for the GP model):
of the trapezoidal 10  6 10  6 10  5 10  4 10  5
λg ∞ approximation
f xGP (t ) = ∫xβ [log (z ) − Γ ′ (λ g t ) /Γ (λ g t ) ] z λg t − 1e−zdz,
Γ (λ g t ) g Fx (Δt , tN ) in (72)
⎡ ∞ ⎛ ⎞ ⎤ 54.99 49.99 59.16 59.50 74.75
(λt )⌊x / y⌋ γ (k , βx ) (λt ) k ⎜ k − 1⎟ − 1⎥,
 [L ]
f xδ (t ) = λe−λt f xExp (t ) = − λe−λt ⎢ ∑ ⎜ λt ⎟
⌊x /y ⌋! ⎢ ⎥ % Error  [L ] 10−6% 10−3% – 10−4% –
⎣ k = 1 (k − 1 ) ! k! ⎝ ⎠ ⎦ (74)
 [L2] 3517 3000 3903 4630 7421
with ⌊·⌋ the integer part function, Γ (x ) the Gamma function and
%Error  [L2] 10−6% 10−3% – 10−5% –
γ (k , βx ) the lower incomplete gamma function. The plots obtained
for the GP, CPP-Delta and CPP-Exp models in Fig. 5(a) match ex-  [L3] 252,490 299,980 321,940 436,910 906,330

actly with those obtained from the expressions (74). Furthermore, %Error  [L3] 10−6% 10−3% – 10−4% –

Table 3 shows errors around 10−3% or less for the lifetime mo-
ments  [Ln] (n = 1, 2, 3) compared with the values computed from

Fig. 5. pdf's fx(t) of the lifetime L of a system with threshold level x ¼99 for several Lévy deterioration models: (a) progressive GP (λg = 0.1, βg = 1/20) and several CPP's
models. (b) Combined degradation: the GP model with each of the CPP's models. The CPP Poisson rate is λ = 0.1 in all cases.
176 J. Riascos-Ochoa et al. / Probabilistic Engineering Mechanics 45 (2016) 164–176

expressions (74). This demonstrates the accuracy of the proposed and performance of structures and infrastructure: a brief review, Struct. In-
numerical method for the reliability estimation of the Lévy de- frastruct. Eng. 8 (1) (2012) 1–25.
[4] J. van Noortwijk, A survey of the application of gamma processes in main-
gradation model. tenance, Reliab. Eng. Syst. Saf. 94 (2009) 2–21.
Finally, Fig. 5(b) shows the pdf's for the combined cases; thus [5] M. Sánchez-Silva, G.-A. Klutke, D.V. Rosowsky, Life-cycle performance of
each CPP model was combined with the progressive gamma de- structures subject to multiple deterioration mechanisms, Struct. Saf. 33 (2011)
206–217.
gradation GP (λg = 0.1, βg = 1/20). It can be observed that the
[6] I. Iervolino, M. Giorgio, E. Chioccarelli, Gamma degradation models for
combined models lead to smaller failure times, which is expected earthquake-resistant structures, Struct. Saf. 45 (2013) 48–58.
since we have added two sources of degradation. [7] U. Sumita, J.G. Shanthikumar, A class of correlated cumulative shock models,
Adv. Appl. Probab. 17 (2) (1983) 347–366.
[8] M.A. Wortman, G.-A. Klutke, H. Ayhan, A maintenance strategy for systems
subjected to deterioration governed by random shocks, IEEE Trans. Reliab. 43
8. Conclusions (3) (1994) 439–445.
[9] T. Nakagawa, Shock and Damage Models in Reliability, Springer, London, 2007.
[10] D. Montoro-Cazorla, R. Pérez-Ocón, A shock and wear system under en-
The framework proposed in this paper, for the modeling and vironmental conditions subject to internal failures, repair, and replacement,
reliability estimation of deteriorating systems, was supported on Reliab. Eng. Syst. Saf. 99 (2012) 55–61.
the assumptions of independent, stationary and non-negative in- [11] I. Iervolino, M. Giorgio, E. Chioccarelli, Closed-form aftershock reliability of
damage-cumulating elastic-perfectly-plastic systems, Earthq. Eng. Struct. Dyn.
crements (Section 4), which allow defining the deterioration pro- 43 (2014) 613–625.
cess as an increasing Lévy process (or Subordinator). These as- [12] J. Riascos-Ochoa, M. Sanchez-Silva, R. Akhavan-Tabatabaei, Reliability analysis
sumptions and specific cases of Lévy deterioration models (e.g., of shock-based deterioration using phase-type distributions, Probab. Eng.
Mech. 38 (2014) 88–101.
CPP models, gamma processes, deterministic drift, and combined
[13] M.D. Pandey, X.X. Yuan, J.M. van Noortwijk, The influence of temporal un-
degradation) have been used previously in several engineering certainty of deterioration on life-cycle management of structures, Struct. In-
applications, e.g., in reliability engineering [1,18,4,9], earthquake frastruct. Eng. 5 (2) (2009) 145–156.
engineering [6,11], operations research [24]. The advantages of our [14] R. Kumar, P. Gardoni, M. Sánchez-Silva, Effect of cumulative seismic damage
and corrosion on the life-cycle cost of reinforced concrete bridges, Earthq. Eng.
framework is that it provides an analytical definition of the de- Struct. Dyn. 38 (2009) 887–905.
gradation mechanisms (shock-based, progressive and combined) [15] M. Sánchez-Silva, G.-A. Klutke, D.V. Rosowsky, Optimisation of the design of
by means of the Lévy measure (finite or infinite) and/or the de- infrastructure components subject to progressive deterioration and extreme
loads, Struct. Infrastruct. Eng. 8 (2010) 655–667.
terministic drift; it can combine any of these mechanisms under
[16] M. Abdel-Hameed, Life distribution properties of devices subject to a Lévy
the same mathematical formalism with the assumption of in- wear process, Math. Oper. Res. 9 (4) (1984) 606–614.
dependence among them; it provides analytical expressions for [17] M. Abdel-Hameed, Lévy Processes and Their Applications in Reliability and
the mean and central moments of the Lévy deterioration that al- Storage, Springer, New York, 2014.
[18] Y. Yang, G.-A. Klutke, Lifetime-characteristics and inspection-schemes for Lévy
low to compare different Lévy models and may provide greater degradation processes, IEEE Trans. Reliab. 49 (4) (2000) 377–382.
flexibility in fitting to deterioration data (i.e., the method of mixed [19] R. Durret, Probability: Theory and Examples, Cambridge University Press, USA,
Lévy deterioration in Section 5.3); and finally, the framework 2010.
[20] J. Bertoin, Lévy Processes, Cambridge University Press, Cambridge, U.K., 1996.
provides accurate and efficient numerical methods for the relia- [21] K.-I. Sato, Lévy Processes and Infinitely Divisible Distributions, Cambridge
bility evaluation (reliability function, lifetime density and its mo- University Press, Cambridge, 1999.
ments) and sample-paths simulation. The framework applies [22] P.E. Protter, Stochastic Integration and Differential Equations, Springer, Ger-
many, 2004.
special quantities like the Lévy measure and the characteristic
[23] S. Resnick, A Probability Path, Birkhauser, Boston, 1999.
exponent and function, and requires their closed-form expressions [24] J.M. van Noortwijk, R.M. Cooke, M. Kok, A Bayesian failure model based on
for the reliability evaluation. isotropic deterioration, Eur. J. Oper. Res. 82 (1995) 270–282.
[25] M. Giorgio, M. Guida, G. Pulcini, An age- and state-dependent Markov model
Further research in this topic should be directed towards the
for degradation processes, IIE Trans. 120 (2013) 72–79.
inclusion of other progressive degradation models into this fra- [26] M. Guida, M. Giorgio, The inverse gamma process: a family of continuous
mework (e.g., stable processes, inverse gaussian processes). Also, it stochastic models for describing state-dependent deterioration phenomena,
is necessary to extend the present framework to non-stationary Reliab. Eng. Syst. Saf. 43 (2011) 621–632.
[27] S. Ross, Introduction of Probability Models, Academic Press, San Diego, Cali-
processes, in order to reproduce non-linear deterioration trends, fornia, 2007.
and the possible extension to model dependency between the [28] G. Latouche, V. Ramaswami, Introduction to Matrix Analytic Methods in Sto-
different processes. Finally, it is important to formalize the sta- chastic Modeling, Society for Industrial and Applied Mathematics, Philadel-
phia, 1999.
tistical estimation of the parameters of a general Lévy degradation
[29] P. Holgate, The Lognormal Characteristic Function, Commun. Statist.-Theory
model to fit to real data. The method proposed of mixed Lévy Methods 18 (12) (1989) 4539–4548.
processes and the feasible moments analysis could be helpful in [30] M.A. Johnson, M.R. Taaffe, Matching moments to phase distributions: mixtures
this aspect. of Erlang distributions of common order, Commun. Statist. – Stoch. Models 5
(4) (1989) 711–743.
[31] J. Gil-Pelaez, Note on the inversion theorem, Biom. Trust 38 (3/4) (1951)
481–482.
Acknowledgments [32] H. Bohman, Numerical inversion of a characteristic function, Scand. Actuar. J.
(1975) 121–124.
[33] L. Feng, X. Lin, Inverting analytic characteristic functions and financial appli-
This research was funded by the National Department of Sci- cations, Scand. Actuar. J. 4 (1) (2013) 372–398.
ence, Technology and Innovation of Colombia (Colciencias). The [34] L.A. Waller, B.W. Turnbull, J.M. Hardin, Obtaining distribution functions by
numerical inversion of characteristic functions with applications, Am. Stat. 49
first author also acknowledges the Basic Sciences Department of (4) (1995) 346–350.
Universidad de Bogotá Jorge Tadeo Lozano for their support. [35] R.B. Davies, Numerical inversion of a characteristic function, Biom. Trust 60 (2)
(1973) 415–417.

References

[1] G.-A. Klutke, Y. Yang, The availability of inspected systems subject to shocks
and graceful degradation, IEEE Trans. Reliab. 51 (3) (2002) 371–374.
[2] D. Frangopol, M.-J. Kallen, J. van Noortwijk, Probabilistic models for life-cycle
performance of deteriorating structures: review and future directions, Prog.
Struct. Eng. Mater. 6 (4) (2004) 197–212.
[3] D. Frangopol, D. Saydam, S. Kim, Maintenance, management, life-cycle design