Levy,´ Non-Gaussian Ornstein-Uhlenbeck, and Markov Additive
Processes in Reliability Analysis
A Dissertation
Presented to
the Faculty of the Department of Industrial Engineering
University of Houston
In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy
in Industrial Engineering
By
Yin Shu
August 2016 Levy,´ Non-Gaussian Ornstein-Uhlenbeck, and Markov Additive
Processes in Reliability Analysis
Yin Shu
Approved: Chair of the Committee Qianmei Feng, Associate Professor, Department of Industrial Engineering
Co-Chair of the Committee Hao Liu, Associate Professor, Duncan Cancer Center-Biostatistics Baylor College of Medicine
Committee Members:
Gino Lim, Professor, Department of Industrial Engineering
Jiming Peng, Associate Professor, Department of Industrial Engineering
Edward P.C. Kao, Professor, Department of Mathematics
Suresh K. Khator, Associate Dean, Gino Lim, Professor and Chair, Cullen College of Engineering Department of Industrial Engineering Acknowledgments
First and foremost I would like to express my deepest gratitude to my advisor Prof.
Qianmei Feng. The interesting and beautiful topic proposed by her made this dissertation possible. And it is my great honor that Prof. Hao Liu (Duncan Cancer Center-Biostatistics,
Baylor College of Medicine) has accepted to be my co-advisor. Here I appreciate all their advising, support, encouragement and help to make my Ph.D. research and life stimulating.
I am lucky to be one of their students.
I am especially grateful to Prof. Edward P.C. Kao for his excellent courses and exper- tise in stochastic processes. I also would like to thank Prof. David Coit for his professional comments on my research work. Many thanks to Prof. Gino Lim and Prof. Jiming Peng for serving as my committee members, and for their time, interest, and insightful suggestions.
I also would like to thank all the faculties, staff, and friends who have contributed immensely to my personal and professional time at the University of Houston. I acknowl- edge the financial support from Texas Norman Hackerman Advanced Research Program under Grant no.003652-0122-2009 and USA National Science Foundation under Grant no.0970140.
Most of all, I thank my parents and love for their understanding and encouragement.
Their support made me devote my time in completing my Ph.D. study. This dissertation is dedicated to them.
iv Levy,´ Non-Gaussian Ornstein-Uhlenbeck, and Markov Additive
Processes in Reliability Analysis
An Abstract
of a
Dissertation
Presented to
the Faculty of the Department of Industrial Engineering
University of Houston
In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy
in Industrial Engineering
By
Yin Shu
August 2016
v Abstract
Unavoidable degradation is one of the major failure mechanisms of many systems
due to internal properties (mechanical, thermal, electrical, or chemical) and external in-
fluences (temperature, humidity, or vibration). Such degradation in critical engineering
systems (e.g., pipelines, wind turbines, power/smart grids, and mechanical devices, etc.)
takes the form of corrosion, erosion, fatigue crack, deterioration or wear that may lead to
the loss of structural integrity and catastrophic failure. Therefore, developing stochastic
degradation models based on appropriate stochastic processes becomes imperative in the
reliability and statistics research communities.
This dissertation aims to develop a new research framework to integrally handle the
complexities in degradation processes (the intrinsic/extrinsic stochastic properties, com-
plex jump mechanisms and dependence) based on general stochastic processes including
Levy,´ non-Gaussian Ornstein-Uhlenbeck (OU), and Markov additive processes; and to
develop a new systematic methodology for reliability analysis that provides compact and
explicit results for reliability function and lifetime characteristics. First, to handle the
intrinsic stochastic properties and complex jumps, we use Levy´ subordinators and their
functional extensions, Levy´ driven non-Gaussian OU processes, to model the cumula- tive degradation with random jumps. We then integrally handle the complexities of a degradation process including both intrinsic and extrinsic stochastic properties with com- plex jump mechanisms, by constructing general Markov additive processes. Moreover, the models are extended to multi-dimensional cases for multiple dependent degradation processes under dynamic environments, where the Levy´ copulas are studied to construct
vi Markov-modulated multi-dimensional Levy´ processes. The Fokker-Planck equations for such general stochastic processes are developed, based on which we derive the explicit results for reliability function and lifetime moments, represented by the Levy´ measures,
the infinitesimal generator matrices and the Levy´ copulas. To analyze the degradation data
series from such degradation phenomena of interest, we propose a systematic statistical
estimation method using linear programming estimators and empirical characteristic func-
tions. We also construct bootstrap procedures for the confidence intervals. Simulation
studies for Levy´ measures of gamma, compound Poisson, positive stable and positive
tempered stable processes are performed. The framework can be recognized as a general
approach that can be used to flexibly handle stylized features of widespread classes of
degradation data series such as jumps, linearity/nonlinearity, symmetry/asymmetry, and
light/heavy tails, etc. The results are expected to provide accurate reliability prediction
and estimation that can be used to assist the mitigation of risk and property loss associated
with system failures.
vii Table of Contents
Acknowledgments ...... iv
Abstract ...... vi
Table of Contents ...... viii
List of Figures ...... xii
List of Tables ...... xv
1 Introduction ...... 1
1.1 Background & Motivation ...... 4
1.2 Problem Statement ...... 7
1.3 Objectives & Contributions ...... 10
1.4 Organization ...... 13
2 Literature Review ...... 15
2.1 Wiener Processes ...... 15
2.2 Compound Poisson Processes ...... 18
2.3 Gamma Processes ...... 21
2.4 Levy´ Processes ...... 23
viii 2.5 Ornstein-Uhlenbeck Processes ...... 25
2.6 Markov Additive Processes ...... 28
3 Life Distribution Analysis Based on Levy´ Subordinators for Degradation
with Random Jumps ...... 31
3.1 Introduction ...... 32
3.2 Preliminaries of Levy´ Processes ...... 35
3.2.1 Characteristics ...... 36
3.2.2 Special Cases of Levy´ Processes ...... 39
3.3 Life Distribution Analysis Based on Levy´ Subordinators ...... 41
3.4 Life Distribution Analysis for Temporally Homogeneous Gamma Process
with Random Jumps ...... 45
3.4.1 Reliability Function Using Traditional Convolution Approach . . 46
3.4.2 Reliability Function Using Levy´ Measures ...... 47
3.5 Numerical Examples ...... 51
3.6 Conclusions ...... 54
4 Levy´ Driven Non-Gaussian Ornstein-Uhlenbeck Processes for Degrada-
tion-based Reliability Analysis ...... 56
4.1 Introduction ...... 56
4.2 Preliminaries ...... 61
4.2.1 Levy-It´ oˆ Decomposition ...... 61
4.2.2 Model Construction ...... 62
ix 4.3 Reliability Function and Lifetime Moments ...... 64
4.3.1 Results Based on Levy´ Subordinators ...... 67
4.3.2 Results Based on Non-Gaussian OU Processes ...... 71
4.4 Numerical Examples ...... 75
4.5 Conclusions ...... 79
5 Markov Additive Processes for Degradation with Jumps under Dynamic
Environments ...... 81
5.1 Introduction ...... 82
5.2 Model Construction ...... 85
5.3 Fokker-Planck Equations for Markov Additive Processes ...... 87
5.4 Reliability Function and Lifetime Moments ...... 94
5.5 Numerical Examples ...... 100
5.6 Conclusions ...... 107
6 Markov-modulated Multi-dimensional Levy´ Processes for Multiple Depen-
dent Degradation Processes under Dynamic Environments ...... 108
6.1 Introduction ...... 108
6.2 Preliminaries ...... 111
6.2.1 Multi-dimensional Levy´ Processes ...... 111
6.2.2 Model Construction ...... 114
6.3 Fokker-Planck Equations ...... 118
6.4 Reliability Function and Lifetime Moments ...... 120
x 6.5 Numerical Examples ...... 124
6.6 Conclusions ...... 129
7 Statistical Inference ...... 130
7.1 Introduction ...... 130
7.2 Estimation for Levy´ Degradation Processes ...... 132
7.2.1 Point Estimates ...... 132
7.2.2 Bootstrap Confidence Intervals ...... 134
7.3 Estimation for OU Degradation Processes ...... 136
7.3.1 Point Estimates ...... 136
7.3.2 Bootstrap Confidence Intervals ...... 139
7.4 Simulation Study ...... 140
7.5 Case Study ...... 146
7.6 Conclusions ...... 150
8 Summary and Discussions ...... 151
References ...... 155
Appendices ...... 181
xi List of Figures
Figure 1.1 Complexities in degradation ...... 7
Figure 3.1 Reliability function for gamma degradation with additional gamma
jumps ...... 52
Figure 3.2 Reliability function for gamma degradation with three jump types . 52
Figure 3.3 Pdf of lifetime for gamma degradation with three jump types . . . 53
Figure 3.4 Hazard rate for gamma degradation with three jump types . . . . . 53
Figure 4.1 Sample paths of Levy´ processes ...... 59
Figure 4.2 Sample paths of non-Gaussian OU processes ...... 60
Figure 4.3 Reliability function w.r.t. time t and failure threshold x based on
˜ Xs(t) ...... 77
Figure 4.4 Reliability function w.r.t. time t and failure threshold y based on Y (t) 77
˜ Figure 4.5 Reliability function w.r.t. time t based on Xs(t) and Y (t) . . . . . 78
˜ Figure 4.6 First moments of lifetime w.r.t. failure threshold based on Xs(t)
and Y (t) ...... 78
Figure 5.1 A sample path of Markov additive process with random jumps
when the environment states change ...... 87
xii Figure 5.2 A sample path of Markov additive process with no jump when the
environment states change ...... 93
Figure 5.3 Reliability function w.r.t. time t and failure threshold x for Case 1 . 104
Figure 5.4 Reliability function w.r.t. time t and failure threshold x for Case 2 . 104
Figure 5.5 Reliability functions w.r.t. time t when x = 15 and x = 20 for both
Case 1 and Case 2 ...... 105
Figure 5.6 First moments of lifetime w.r.t. failure threshold for both Case 1
and Case 2 ...... 105
Figure 5.7 Second moments of lifetime w.r.t. failure threshold for both Case 1
and Case 2 ...... 106
Figure 6.1 A sample path of Markov-modulated two-dimensional Levy´ pro-
cess with random jumps when the environment states change ...... 117
Figure 6.2 The mean of the first passage time w.r.t. failure threshold for Case 1 128
Figure 6.3 The mean of the first passage time w.r.t. failure threshold for Case 2 129
Figure 7.1 Levy´ density of PS(κ); dashed: estimated when n=25; solid: true . 142
Figure 7.2 Increments of the background driving PS(κ); dashed: estimated
when n=50; solid: true ...... 144
Figure 7.3 Marginal Levy´ density of OU driven by PS(κ); dashed: estimated
when n=50; solid: true ...... 145
˜ Figure 7.4 90% confidence intervals of reliability function for Xs(t) . . . . . 145
Figure 7.5 90% confidence intervals of reliability function for Y (t) ...... 146
xiii Figure 7.6 The capacity losing processes of four 18650 Li-ion batteries . . . . 147
Figure 7.7 Estimation of reliability function when failure threshold is 0.9 . . . 148
Figure 7.8 Estimation of first moments of lifetime w.r.t. failure threshold . . . 149
Figure 7.9 Estimation of standard deviation of lifetime w.r.t. failure threshold 149
xiv List of Tables
Table 3.1 Parameter values for models in Chapter 3 ...... 51
Table 4.1 Parameter values for models in Chapter 4 ...... 76
Table 5.1 Parameter values for models in Chapter 5 ...... 103
Table 6.1 Parameter values for models in Chapter 6 ...... 128
Table 7.1 Results of κˆ for PS(κ) ...... 142
Table 7.2 Results of αˆ and κˆ for OU driven by PS(κ) ...... 144
xv Chapter 1
Introduction
The research conducted in this dissertation focuses on exploring Levy´ processes, non-Gaussian Ornstein-Uhlenbeck (OU) processes, and Markov additive processes for modeling complex stochastic degradation processes, with the aim of developing systematic procedures for deriving/evaluating reliability and lifetime characteristics. The results are expected to provide precise reliability prediction and estimation that can be used to assist the mitigation of risk and property loss associated with system failures.
With the advancement in technology, new and complex engineering systems (e.g., wind turbines, power/smart grids, subsea pipelines, and mechanical devices, etc.) have increasingly high reliability. However, they usually experience degradation processes by gradually losing their intended functionality over time. The degradation takes forms of wear, fatigue, erosion, corrosion and aging, etc. [3, 75, 181, 206]. A system fails when the accumulated degradation (e.g., fatigue crack) hits a boundary. Degradation data series typically contain more life-related information than the traditional failure time data, es- pecially for highly-reliable systems. Therefore, developing stochastic degradation models based on appropriate stochastic processes becomes increasingly critical and significant in the research community.
Levy´ processes are a class of cadl` ag` (right continuous with left limits) Markov pro- cesses with independent and stationary increments [17, 168]. The general Levy´ process
1 mathematically consists of three independent parts: the deterministic linear drift part, the
Wiener/Brownian part, and the pure jump part. As stochastic processes with random jump-
s, Levy´ processes have been widely used in modeling stochastic fluctuations for mathe-
matical finance [55, 170] and risk management [118, 119, 124]. However, they have not
been well studied for degradation modeling (e.g., no explicit results of life characteristics
based on the general Levy´ degradation process). One of the most important advantages of
using Levy´ processes to model degradation is that their jump parts represented by Levy´ measures can model a great deal of jump mechanisms in degradation. We explore Levy´ processes for analyzing life distribution and reliability characteristics in Chapter 3 and a part of Chapter 4.
OU processes, another class of continuous time continuous state stochastic process- es, were named after L.S. Ornstein and G. E. Uhlenbeck [194] in a physical modeling context, where the background driving process is a Wiener process, thus called ordinary or Gaussian OU processes [138]. Non-Gaussian OU processes are the generalization of ordinary OU processes by replacing Wiener processes with non-Gaussian Levy´ processes
(e.g., positive tempered stable processes). They have been recently developed and applied in financial models by [21–23]. To our best knowledge, non-Gaussian OU processes have not been used for degradation modeling. In fact, it is nontrivial to obtain a closed-form probability distribution function for an OU process driven by a Levy´ process. The most important advantage of using non-Gaussian OU processes in modeling degradation pro- cesses stems from their flexibility in modeling stylized features of degradation data series such as jumps, asymmetry, and heavy tails. We explore non-Gaussian OU processes for
2 studying reliability as part of Chapter 4.
Markov additive processes are a class of binary stochastic processes with one compo- nent as an additive process (e.g., Levy´ process) that is modulated by the other component, which is a standard Markov process [42, 46]. They form one of the most popular cases of Markov-modulated processes, and have been applied in queueing and storage systems
[151]. However, Markov additive processes have not been well studied for degradation analysis. Since Markov additive processes can represent the dependence of degradation on external factors (e.g., environments), they are suitable for modeling degradation under the dynamic (time-varying) environment. We explore this superiority in Chapter 5.
Multi-dimensional Levy´ processes can be constructed using Levy´ copulas [55, 109].
The dependence among components of a multi-dimensional Levy´ process can be com- pletely characterized by a Levy´ copula, a function that has the similar properties with the ordinary copula but is defined on a different domain. Markov-modulated multi- dimensional Levy´ processes are multi-dimensional cases of Markov additive processes by setting the additive component to be a multi-dimensional Levy´ process. They are appropriate candidates to model multiple dependent degradation processes. We explore this potential in Chapter 6.
Traditional maximum likelihood estimation and Bayesian estimation are not conve- nient for such general jump processes without closed-form distributions. [55] provided a highly comprehensive and thorough treatment of Levy´ processes in finance, covering Levy´ models, simulation and estimation. [21] showed that it is not straightforward to implement traditional likelihood-based estimation procedures for the non-Gaussian OU-based model,
3 although various moment-based methods are simple to use. [107] proposed the cumu- lant M-estimator (CME) to estimate the parameters in Levy´ processes. We explore this approach to estimate/predict the reliability characteristics in Chapter 7 using degradation data series.
1.1 Background & Motivation
Reliability of systems is one of the most critical concerns in many fields including energy, health, aerospace, and national defense, etc. In investigating reliability, degra- dation processes have been considered and analyzed for more than half a century. S- ince the 1970s, stochastic processes have been used to represent degradation evolution in order to handle the stochastic properties in degradation processes. Erhan C¸inlar, Mo- hammad Abdel-Hameed, and William Meeker, among others, have made substantial and foundational contributions to the general area of research in degradation-based reliability analysis. [40–53] mainly focused on developing mathematical theories of stochastic pro- cesses (e.g., Markov additive processes, Levy´ systems of Markov additive processes). The work provided the prerequisite mathematical support in degradation modeling using such stochastic processes. [3–15] mainly studied life distribution properties (e.g., increasing failure rate) and the optimal maintenance/inspection policies based on stochastic degrada- tion models. [140–147] developed statistical methods for degradation data analysis. Their research demonstrates that constructing appropriate stochastic processes in degradation modeling is critical in precisely evaluating and predicting reliability characteristics of
4 highly-reliable systems. Therefore, modeling the degradation processes and then deriv-
ing/estimating reliability characteristics have attracted great attention among mathemati-
cians and engineers in recent years. However, the complexities in degradation lead to more
challenging research topics that need to be further explored.
Two complex stochastic properties of degradation processes in investigating relia-
bility are: 1) the internally-induced stochastic evolution that has inherent statistical un- certainties stemming from physical, electrical, thermal, or chemical features of systems, such as molecule/atom structure, arrangement and composition, etc. (e.g., the increments of the wear volume during a fixed time interval follow an inverse Gaussian distribution), and 2) the externally-induced stochastic evolution that stems from dynamic (time-varying) environment factors such as temperature, pressure and humidity, etc. These extrinsic un- certainties in dynamic environments can be described as two different aspects: 1) different evolution patterns of degradation under different environment states, e.g., the degradation process evolves as a positive tempered stable process with drift under one environment state, while it evolves as a compound Poisson process without drift under another state, and 2) an instantaneous random jump occurring in degradation at the time the state of the environment process changes.
In addition, there are complex jump mechanisms embedded in the stochastic evolution of degradation processes. As jumps due to the internal operation (e.g., sudden electrical short, chemical reaction, etc.) under the deterministic environment, internally-induced jumps can be an infinite number of small jumps in any small interval of time, or a finite number of big jumps that occur according to a probability law. As jumps due to the
5 changes in environment states (e.g., temperature extremes, vibration/tension shocks, etc.),
the emergence of externally-induced jumps follows the law of the environment process.
The size of these jumps is a random variable that can be dependent on the switching
environment states. Moreover, multiple degradation processes in a system are naturally
interdependent including internally-induced dependence due to the internal features and
externally-induced dependence due to the exposure to the same external environment con-
ditions.
All the complexities of degradation processes are illustrated in Figure 1.1. The two
stochastic properties are demonstrated by the stochastic evolution of the binary stochas-
tic processes {X(t),E(t)}, where the two-dimensional stochastic degradation process
X(t) = [X1(t),X2(t)] is modulated by the environment process E(t). Under a certain state of E(t), X(t) evolves as a two-dimensional dependent stochastic process. During different states of E(t), X(t) evolves in different patterns with different jump mechanisms.
In addition, instantaneous jumps in all degradation processes induced by the changes of
E(t) are also properly demonstrated.
The motivation for this dissertation is to provide a new methodology in degradation-
based reliability analysis, where the complexities in degradation, including the internally-
/externally-induced uncertainties, dependence and complex jump mechanisms, are inte-
grally modeled in a broad class of general stochastic processes. Based on the mathematical
theories of such stochastic processes that are analytically appealing, we aim to develop
the systematic procedures in deriving/evaluating life characteristics that can provide the
consolidated and compact results.
6 X(t) X1(t)
X2(t)
t i k l m E(t) j
Figure 1.1: Complexities in degradation
1.2 Problem Statement
Depending on the availability of environmental data, two classes of stochastic models have been generally constructed to represent the degradation processes. When it is not possible to identify the factors or it is not feasible/economically convenient to monitor the external environment (in this case, randomness in the environment process and/or its effect on degradation evolution is ignored and the environment is assumed to be deterministic), the observable degradation process was directly represented by a stochastic process mod- eling the intrinsic uncertainties with temporal variability [3, 75, 152, 206] (e.g., Wiener process for non-monotonic increments, gamma process for monotonic increments, and
7 compound Poisson process for pure jump increments); or the unobservable degradation process was treated as a latent process, measured and tracked by internal stochastic covari- ates that are observable marker processes [86, 104, 105, 175, 208] (e.g., Linear processes with random effects, Wiener processes, and Poisson processes). These stochastic models just represented internally-induced uncertainties.
The second class of stochastic models is constructed when the operating environment can be monitored and measured. In [71], a stochastic covariate failure model was studied for assessing system reliability, where the external stochastic covariates were modeled by
Wiener-based diffusion processes. The life distribution was assumed to be explicitly relat- ed to such stochastic covariates. However, the stochastic models just represent externally- induced uncertainties with temporal variability. Some advanced stochastic degradation models were constructed in [112–115], where the degradation process is modulated by the environment process. A compound Poisson process and a linear process were used to represent the conditional degradation process, and the corresponding parameters are explicit functions of the environment process, which is represented by a Markov chain with a finite number of states. These models are motivated by the original ideas from [40], in which Markov additive processes were first proposed in degradation modeling.
Advancement in technology have had and will continue to impel researchers to make improvement in stochastic models for degradation-based reliability analysis. Recently, advanced technology in measurement and deployment of sensors/smart chips make it con- venient to collect plenty of degradation data and environment data, which are accurate, re- liable and real-time. However, the existing stochastic models for representing degradation
8 processes have left a number of practical issues untouched that stem from the complexities in degradation. Such practical problems are itemized as follows.
1. The common linear-, Wiener-, gamma-, inverse-Gaussian- and Poisson-based mod-
els are not flexible in general cases, because it may not be appropriate in reality to
assume that the increments follow normal, inverse-Gaussian or gamma distributions,
or to assume that the jumps occur according to the Poisson laws. In practice, the
degradation data series have complex jump mechanisms, and have asymmetric and
heavy-tail properties when there are big jumps occurring randomly.
2. There are no systematic and extendible procedures in analyzing the degradation
evolution and deriving life characteristics (e.g., reliability function and lifetime mo-
ments) from stochastic degradation processes. Existing approaches are typically
based on the convolution formula, under the condition that the distribution functions
of stochastic processes of interest have closed-forms. They often can not produce
compact and explicit results, e.g., when the degradation with random jumps is mod-
eled by the sum of a stable process and a compound Poisson process.
3. The existing stochastic models for degradation processes under the dynamic envi-
ronment are limited: they are simply based on linear processes and Poisson pro-
cesses; and they did not handle the jumps occurring at the time the state of the
environment process is switching, which is a common phenomenon in degradation.
4. In order to develop general degradation models under the dynamic environmen-
t, it is necessary to describe 1) the general stochastic process for degradation, 2)
9 the general stochastic process for environment, and 3) the effects of environment
on degradation. The development of the characteristics of such processes and the
subsequent reliability analysis based on such characteristics is a nontrivial work,
even for simple cases. The difficulty of these stems from 1) the stochastic evolution
of degradation has complex mechanisms such as random jumps, 2) the stochastic
nature of environment, and 3) the distributional derivation for the first passage time.
5. Stochastic models for multiple dependent degradation processes are not well devel-
oped. The existing models are mainly based on multi-dimensional Wiener processes,
where the dependence is simply represented by the covariance matrix. These models
cannot handle the jumps, the dependence among the jumps, and the dependence
stemming from the environment process.
These issues can be concisely summarized as 1) the lack of general and systemically analyzable stochastic models, 2) the lack of integrated uncertainties, complexities and dependence management methodology, 3) the lack of systematic procedures on derivation, and 4) the lack of explicit and powerful results/formulas.
1.3 Objectives & Contributions
This dissertation aims to develop a new research framework to integrally handle the complexities in degradation processes (the intrinsic and extrinsic stochastic proper- ties, dependence and complex jump mechanisms) based on general stochastic processes including Levy´ processes, non-Gaussian OU processes and Markov additive processes,
10 and to construct a new methodology for reliability analysis, in which all the critical prob-
lems mentioned in the previous section are well addressed. The results provide efficient
and accurate reliability evaluation and prediction. The specific objectives are itemized as
follows.
Objective 1: we develop a class of general stochastic models that can represent the
internally-induced stochastic properties, dependence and complex jumps in degra-
dation processes. We first consider the Levy´ processes for degradation modeling by
relaxing the assumptions on the law of increments and the law of jumps, in which
the jump mechanisms of degradation are well described by the Levy´ measures. We
further extend Levy´ processes to non-Gaussian OU processes that relax the assump-
tion of the linear mean path (Chapters 3, 4). This provides a cornerstone for the next
level work.
Objective 2: we investigate a class of general stochastic models that can simultaneously
represent the internally- and externally-induced stochastic properties, dependence
and complex jumps in degradation processes. We study Markov-modulated pro-
cesses, a class of binary stochastic processes, in which one component is used to
model the degradation evolution and the other is used to model the environment
evolution (Chapter 5). Multi-dimensional setting is further explored using the Levy´
copulas (Chapter 6).
Objective 3: we explore the potentials of the inverse Fourier transform in deriving
11 compact- and closed-form results for reliability characteristics from Levy´ degrada-
tion processes. We extend the idea to be applied in multiple dependent degradation
modeled by multi-dimensional Levy´ processes (Chapter 3).
Objective 4: we propose to use the Fokker-Planck equations of such analytically appeal-
ing stochastic processes (Levy,´ non-Gaussian OU, and Markov-modulated process-
es) in the derivation of reliability characteristics based on the Laplace transforms.
We develop systematic procedures to derive and obtain explicit and compact results
(Chapters 4, 5, 6).
Objective 5: we propose a systematic statistical inference method using linear program-
ing estimators and empirical characteristic functions. The point estimates of re-
liability function and lifetime moments are obtained by deriving their explicit ex-
pressions. Bootstrap procedures are also constructed for the confidence intervals
(Chapters 7).
To achieve these five objectives, several fundamental research areas need to be ex-
plored as part of the research methodology. The development of the integrated method-
ology enhances the multidisciplinary research by expanding mathematics, probability and
statistics, and computer science into reliability engineering. The work will provide accu-
rate and realistic new models for degradation analysis, by realizing multiple uncertainty
sources of degradation mechanisms. The framework can be recognized as a general ap-
proach that can be used to handle widespread classes of degradation data sets, including
linear data, nonlinear data, fluctuant data, etc.
12 1.4 Organization
The dissertation summarizes the results in Shu, et al. [176–180]. Chapter 2 provides
a comprehensive literature review for the related research on degradation modeling, Levy´ processes, non-Gaussian OU processes, and Markov-modulated processes. To handle the intrinsic stochastic properties with complex jumps, Chapter 3 presents the degradation models using Levy´ processes. The complex jump mechanisms including infinite activities
(IA) and finite activities (FA) in degradation are well described by the Levy´ measures.
Based on the inverse Fourier transform, the new closed-forms of reliability function and probability density function of lifetime are derived. In Chapter 4, we further explore the
Levy´ subordinators and extend the model by using non-Gaussian OU processes, which can model the degradation processes with non-linear mean paths. Based on the Fokker-Planck equations, the new explicit results in Laplace expressions for both reliability function and lifetime moments are derived. To handle both the intrinsic and extrinsic stochastic proper- ties with complex jumps, in Chapter 5, we integrally model the complexities of degradation processes under the dynamic environment by constructing Markov additive processes. The
Fokker-Planck equations are developed for such binary processes. The explicit results for reliability function and lifetime moments are derived. Chapter 6 explores the multi- dimensional settings for the stochastic processes of interest based on Levy´ copulas to han- dle the dependence among multiple degradation processes. The Markov-modulated multi- dimensional Levy´ processes are constructed to model the multiple dependent degradation
13 processes under dynamic environments. Chapter 7 provides the systematic estimation pro- cedures for our stochastic models. The simulation and case studies are performed. Finally,
Chapter 8 is devoted to make summary and discuss some potential research directions following this dissertation.
14 Chapter 2
Literature Review
In this chapter, we do a thorough literature review for the stochastic processes relat- ed to the research in this dissertation. Wiener processes, compound Poisson processes, and gamma processes are widely used in degradation-based reliability analysis, since the required mathematical derivations are relatively straightforward based on the exist- ing closed-form distribution functions of these processes. The fundamental mathematical theories of general Levy´ processes, non-Gaussian OU processes, and Markov additive processes are summarized. Some advanced mathematical theories explored by mathe- maticians and their applications in financial economics are enthusiastically studied and reviewed in this chapter.
2.1 Wiener Processes
Wiener processes were named after Norbert Wiener [209], who has made a great con- tribution in the mathematical description of Brownian motion (named after Robert Brown, who described the random motion of particles suspended in a fluid). Wiener processes are a special case of Levy´ processes with stationary and independent increments following a normal distribution. Based on this property, [171] developed some advanced mathemat- ical theories of Wiener processes. They proved that certain classes of random variables
15 associated with Wiener processes are infinitely divisible. They analyzed the infinitesimal structure and developed the theory of stochastic integrals of Wiener processes. Theoretical research on Wiener processes can be found in [62, 172].
In finance, the price of an option is directly related to the prices of its underlying stocks. Wiener processes are well used in describing the logarithm of stock prices under the seminal Black-Scholes option pricing model [27, 148]. Since then, many extended models based on Wiener processes have been developed. [97] studied the pricing of a
European call on an asset that has a stochastic volatility. The option price was determined in a series form for the case in which the stock price was instantaneously uncorrelated with the stochastic volatility. [195] first studied the default measures for individual firms and assessed the effect of default risk on the equity returns based on such an option pricing model, where the logarithm of the market value of a firm’s underlying assets was described by a Wiener process, and the default was defined as the failure to service debt obligations.
Similar studies can be found in [58, 85, 89]. These Wiener-based models have had a profound impact on financial economics.
A Wiener process is appropriate for modeling degradation that changes nonmono- tonically because it can have non-negative and negative increments alternately. To derive the reliability function from a Wiener degradation process, the failure time can be defined as the first passage time to the failure threshold [39, 56, 127]. Using the fact that the
first passage time of the Wiener process follows the inverse Gaussian distribution, Wiener- based degradation models have been well developed for reliability analysis. [206–208] proposed some stochastic degradation models where the underlying process is a Wiener
16 process. In [206], the author considered a Wiener process with normal-distributed mea- surement errors in order to fit the degradation data measured from imperfect instruments, procedures and environments. Inference procedures were constructed based on maximum likelihood estimation and some practical extensions were discussed. In [207], a Wiener process with a time scale transformation was taken into account to model the degradation process. The model and inference methods were illustrated with a case application involv- ing self-regulating heating cables. This model did not consider measurement errors. In
[208], a degradation model with stochastic time-varying covariates was constructed. The degradation is called a latent process that is unobservable, and the stochastic covariate is called a marker process that is observable [108]. These two processes were represented by a bivariate Wiener process, based on which the failure time distribution was derived.
Several candidate markers were discussed for extensions.
Many extensions have been done based on Wiener degradation processes. [69] con- sidered there are many cracks evolving simultaneously in a system, and proposed models to describe the stochastic behavior of fatigue cracks. In the models the growth rates of cracks were described by Wiener processes. The distribution of failure time and its properties were studied, and the parameters were estimated by the maximum likelihood method. [183, 184] presented statistical methods for the remaining useful life estimation based on Wiener processes. A recursive filter algorithm and Bayesian updating frame- work were studied. The updating is done at the time that a new piece of degradation data becomes available. Gyros in an inertial navigation were used as a practical case study to demonstrate the superiority of the proposed model. [152, 153] developed degradation
17 models with several accelerating variables for statistical inference based on both observed
failure values and degradation measurements. The underlying degradation process that
they assumed was a geometric Brownian motion (the logarithm of the degradation process
was assumed to be a Wiener process). The inverse Gaussian distribution provided a good
approximation to the failure distribution. Many other Wiener-based degradation analysis
can be found in [64, 188, 192].
2.2 Compound Poisson Processes
Poisson processes were named after Simeon´ Denis Poisson [161]. They are a class of counting processes and their standard definitions can be found in [167]. Compound
Poisson processes are a generalisation of Poisson processes and are a special case of Levy´ processes with stationary and independent increments evolving by pure jumps. The jumps occurrence is following a Poisson process and their sizes are independent and identically distributed (i.i.d.) random variables. The mathematical theories of compound Poisson processes including Poisson random measures, corresponding Levy´ measures, and charac-
teristic functions are introduced in [168]. Due to these available theories, they are popular
in modeling random events that occur at random discrete times.
In biology, [99] introduced a compound Poisson process to model the rate variation
across lineages on a tree, where the events of substitution rate change were placed onto
a phylogenetic tree according to a Poisson process. The parameters of the model were
estimated based on the Metropolis-Hastings-Green algorithm. [116] proposed a compound
18 Poisson process to describe the neutral changes of gene expression over the evolutionary time, where the occurrence of mutations on the DNA level follows a Poisson process and the effects of a mutation on a gene’s expression are i.i.d. random variables. The observable difference in the expression of a gene between two samples was generated by the difference of two independent compound Poisson processes.
In finance, compound Poisson processes were first used in the classical Cramer-
Lundberg model [57] in risk analysis, where the number of claims follows a Poisson process, and the claim sizes are i.i.d. random variables. The ruin probability can be studied based on the well-known Pollaczek-Khinchine formula. [18, 19] studied the ruin probabilities in compound Poisson models, in which the bounds and asymptotic formulas of the ruin probability were obtained from the n-fold convolution, under specific condi- tions on the claim sizes (sub-exponentially distributed). [67, 199] extended the classical compound Poisson risk process to the one perturbed by a Wiener process, and obtained the Lundberg inequality and the asymptotic formula of the ruin probability by the renewal theory. The convolution formula for the probability of ruin was derived when the distri- butions of the individual claim amounts are subexponential as well as exponential, or a combination of exponential distributions. [213] did a further step by considering a con- stant force of interest to invest the surplus, under these perturbed compound Poisson risk processes. They investigated the asymptotic behavior and bounds of the ruin probability as the initial reserve goes to infinity. They also obtained a Laplace expression of the ultimate ruin probability. [191] obtained optimal dividend strategies considering the compound
Poisson insurance risk model when strictly positive transaction costs are included. They
19 constructed a numerical procedure to deal with general distributions of claim sizes.
The estimation of the ruin probability relies on the estimation of the underlying compound Poisson processes. This research has been studied by [73, 74]. Several effi- cient estimation methods for compound Poisson processes have been developed. [34] first introduced the decompounding methods to estimate the parameters of a compound Pois- son process. They studied statistical properties such as the asymptotic normality and the asymptotic validity of bootstrap confidence regions of the estimator. [54, 68] investigated the nonparametric estimation of the jump density of a compound Poisson process from the discrete observation of one trajectory. The estimation procedure was based on the explicit inversion of the operator giving that the law of the increments is a nonlinear transformation of the jump density.
For systems subject to sporadic jump degradation, a compound Poisson process is one of the appropriate candidates to model the cumulative degradation. [75] considered a device subject to a sequence of shocks occurring randomly according to a Poisson process, where a compound Poisson process is proposed to model the accumulated damage. The reliability function is expressed based on the convolution formula, and the increasing haz- ard rate property is studied. The same model was discussed to describe the fatigue damage of materials in [186]. [204, 205] modeled the degradation process with non-decreasing paths by a generalization of Poisson process, which is a marked point process. The degra- dation process was assumed to be generated by a position-dependent marking of a doubly stochastic Poisson process, where the jump intensity depends on a non-negative random variable, compared to a compound Poisson process. They discussed the characteristics of
20 such processes and derived reliability functions for some special cases. [4, 91] introduced the life distribution and its properties for systems subject to pure jump damage processes.
2.3 Gamma Processes
Gamma processes are a special case of Levy´ processes with stationary and inde- pendent increments following a gamma distribution. They are monotonically increasing because of the positivity of the gamma variable. The Levy´ measure and characteristics of gamma processes were introduced in [168]. Some remarkable mathematical properties such as quasi-invariance of gamma processes were introduced in [193, 212]. They showed that gamma processes are renormalized limits of stable processes and emphasized the deep similarity between the gamma process and the Brownian motion.
In finance, [111] proposed to use gamma processes to describe the jumps of asset price. Based on the Malliavin calculus for jump processes, they derived the Greeks for- mulas for derivative securities with both continuous and discontinuous European payoff structures, by making use of a scaling property of gamma processes with respect to the
Girsanov transform. [66] considered to use gamma processes to model aggregate claims in risk theory. They discussed Bayesian estimation for parameters of gamma processes and presented a method for simulating gamma processes. The values of ruin probability including lower and upper bounds were tabulated.
The gamma process is suitable for modeling degradation that progresses in one di- rection by a sequence of small increments. [3] was the first study to use a gamma process
21 for modelling a wear process, in which life distribution properties such as the increasing failure rate were discussed. [125] later presented a gamma model incorporating a random effect for degradation. They used the model to fit the crack-growth data in [102] and presented goodness-of-fit tests (the data was used in [135]). The calculations for fail- ure time distributions were also presented. [20] studied the gamma degradation models with time-dependent covariates. The cases of parameterized and completely unknown mean degradation functions were investigated, and the maximum likelihood estimator of reliability and degradation characteristics were presented based on the degradation data measured without errors.
[152, 153] constructed a new degradation model by incorporating an accelerated test variable based on gamma processes. Both observed failures and degradation measures were used for parameter inference of the life distribution based on the maximum like- lihood estimation, which is illustrated by using real data for carbon-film resistors and fatigue crack size. [132] proposed a maintenance policy for gamma degrading systems.
[29] performed the parametric inference in a perturbed gamma degradation process. De- tailed discussions were given in [196] that provided an overview and survey for applying gamma processes to model degradation under the maintenance context. The methods for estimation, approximation, and simulation of gamma processes were also reviewed. [70] suggested a non-stationary stochastic process to model the wear and derived the reliability function, where the underlying process is a gamma process.
22 2.4 Levy´ Processes
Levy´ processes were named in honor of the French mathematician Paul Levy.´ They form a rich class of Markov processes with independent and stationary increments. Their basic mathematical theories were established in the 1930s. Since then, a large number of new theories as well as novel applications in such diverse areas as physics, econometrics and mathematical finance, have been developed.
The mathematical theories of Levy´ processes have been well developed in many books, monograph and papers. [168] analytically elaborated both basic and advanced knowledge of Levy´ processes by connecting with infinitely divisible distributions. The basic part of the monograph covers characteristic functions, special cases and distribution- al properties; and the advanced part of the monograph covers the Levy-Itoˆ decomposition and Wiener-Hopf factorizations. [17] introduced both the classical theory and stochastic analysis related to Levy´ processes. Basic ideas behind stochastic calculus such as mar- tingales, stopping times, change of measures and stochastic integration were described systemically. [65, 124] considered the fluctuation theory of Levy´ processes. The results focused on the case of Levy´ processes with jumps in only one direction, including sub- ordinators and spectrally negative processes, for which recent theoretical advances have yielded a high degree of mathematical tractability.
There are intense activities focusing on the application of Levy´ processes in financial economics, especially on the option pricing. Levy´ models relax the restrictive assumptions of the Black-Scholes-Merton model by allowing jumps in the underlying asset price and
23 have become popular. [36] considered the problem of pricing contingent claims on a stock
whose price process is modelled by a geometric Levy´ process, in exact analogy with the ubiquitous geometric Brownian motion model. For a class of options that allow early exercise (Bermudan options), [81, 82, 134] used the Levy´ process to model the log-return
process of the asset that is underlying the option on a finite time horizon, and developed
the valuation of the Bermudan option by inverse Fourier and Hilbert transforms. The same
models based on Levy´ processes for pricing options can be referred to [77].
The total revenue of an insurance company is a deterministic increasing process cor-
responding to the accumulation of premiums minus a stochastic process representing the
offset from independent claims sequentially through time. [118, 119] used Levy´ processes
without positive jumps to model insurance risk processes, in which the jumps represent
claim payments, and they studied the properties of the ruin probability, which corresponds
to the distribution of the first passage time of the risk process above a specified bound.
They did not obtain the explicit results for the first passage time. [100, 101] studied a
general perturbed risk process, where the cumulative claims were modelled by a subordi-
nator with finite expectation, and the perturbation was modeled by a spectrally negative
Levy´ process with zero expectation. They derived the survival function for this kind of
risk processes. The calculation of ruin probability for the general case is nontrivial, which
involves the n-fold convolution and trickish integrals.
The application of Levy´ processes to finance has continued to be a highly active and
fast moving area. [55] provided a monograph that is a highly comprehensive and thorough
treatment of Levy´ processes in finance, covering Levy´ models, simulation and estimation.
24 They studied and introduced the significance of using Levy´ processes in modeling financial
time series compared to Brownian motion-based models. A helpful introduction of Levy´ processes in finance, typically aimed at pricing financial derivatives, can be comprehended in [170].
Levy´ processes are a class of cadl` ag` (right continuous with left limits) homogeneous
Markov processes. They are suitable to model degradation processes with random jumps.
A limited number of studies have used general Levy´ processes to model the degradation, and a closed-form life distribution is obtained only when considering the special cases of
Levy´ processes. [5] used Levy´ processes to model the wear and studied its life distribution properties where the threshold is assumed to be random. [211] used special cases of Levy´ processes to model the degradation and jump damages: a gamma process for wear and a compound Poisson process for random jump damages, respectively. They assumed that the threshold is exponentially distributed, which leads to a closed-form lifetime distribution.
[154] derived the distributions of the first-passage time and the last-passage time based on Levy´ degradation processes. [128] considered the class of degradation-threshold-shock models where the degradation is modeled by a process with independent and stationary increments.
2.5 Ornstein-Uhlenbeck Processes
OU processes were named after Leonard Salomon Ornstein and George Eugene Uh- lenbeck [194] in a physical modelling context, where the background driving process is
25 a Wiener process, thus called ordinary or Gaussian OU processes [138]. In the field of physics, the ordinary OU process is represented by the classic Klein-Kramers dynamics
[123]. In the field of finance, the ordinary OU process is well known as the Vasicek model
[198], in which the interest rate was modeled by such a process. The same interest rate model was studied in [36], where the parameters of ordinary OU process were estimat- ed based on the generalized method of moments. [90] used the ordinary OU process to model the instantaneous net yield of oil in a two-factor model for pricing financial and real assets that are contingent on the price of oil. [26] proposed to use the ordinary OU process to model the noise in the field of metrology, which is important in evaluating the measurement system capabilities. In survival analysis, [1] considered a model where the individual hazard rate is a squared function of an ordinary OU process and studied the survival distributions. [160] systemically introduced the probability theories of ordinary
OU processes including Markov properties, stochastic integrals and martingales.
Non-Gaussian OU processes are generalisation of ordinary OU processes by replac- ing the background driving Wiener process with the non-Gaussian Levy´ process (Levy´ process without Gaussian part). [21–23] presented the most important research contribu- tions in the field of non-Gaussian OU processes for financial models during the recent two decades. In [21], they proposed to use a non-Gaussian OU process to model the stochastic volatility in the stochastic differential equations of the log-price of stocks and the log-exchange rates. The background driving non-Gaussian Levy´ process has no neg- ative increments (e.g., Levy´ subordinators). They showed that it is not straightforward to
26 implement the traditional likelihood-based estimation procedures for the proposed mod- els, although various moment-based methods are simple to use. In [22], they analyzed the probability properties of the realized volatility in the context of the non-Gaussian OU- based model, which provided simple quasi-likelihood results that could be used to perform a computationally simple estimation. In [23], they studied the probability properties of the integrated non-Gaussian OU process, which is an important component in the expression of the log-price of stocks or the log-exchange rates, and described the integrated vari- ance in the proposed stochastic volatility models. They explored the tail behavior of the integrated OU process that reflects the tail behavior of the return of underlying stocks.
The efficient statistical inference includes the estimation of non-Gaussian OU pro- cesses is significant in applying such related models in financial data analysis. [162] proposed an indirect inference method for a class of stochastic volatility models for fi- nancial data based on non-Gaussian OU processes. They combined a quasi-likelihood estimator derived from maximizing an approximative Gaussian quasi-likelihood function with simulations, and then applied a method of moments to obtain the indirect estimator, which is better than the pure quasi-likelihood estimator. [189] developed an efficient and explicit estimation procedure for non-Gaussian OU processes based on their characteristic functions, assuming that the marginal law belongs to a parametric family indexed by a parameter vector. The approach can deal with a general case of processes having both positive and negative jumps. The peculiar form of the characteristic functions of non-
Gaussian OU processes and its relation with the characteristic functions of the underlying
Levy´ process were exploited in [190]. Based on the inversion of characteristic function,
27 they provided fast and reliable simulation procedures for OU processes. Simulation based estimation procedures for non-Gaussian OU processes were discussed in [88, 93, 166] .
The approximate results were often implemented since it is difficult to accurately simulate the jumps in the corresponding Levy´ processes.
2.6 Markov Additive Processes
Markov additive processes are a broad class of Markov-modulated processes. Their general definitions and fundamental mathematical theories were introduced by [42, 46].
Since then, many advanced probability theories have been developed. The first passage time is defined as the first time a stochastic process hits a bound. [59] studied the law of the first passage time of a Markov additive process, where the additive component is a spectrally negative Levy´ process (all jumps in the Levy´ process are negative). Based on the theory of Jordan chains, they characterized the law by a matrix function. To express such matrix, a matrix equation needs to be solved, which is intractable. [149] studied the
Wiener-Hopf factorization for a Markov additive process with the multi-dimensional addi- tive component and its generating multi-dimensional reflected process (the reflection takes place when the additive component hits the value zero) using the matrix analytic approach, and then derived a closed-form formula for the stationary distribution of this reflected pro- cess. [103] analyzed the matrix exponent of a Markov additive process with nonnegative jumps, focusing on the roots of the generalized Cramer-Lundberg´ equation that plays an important role in the fluctuation theory such as the first passage time. The same topic of
28 special cases of Markov additive processes such as Markov-modulated Brownian motion
and Markov-modulated compound Poisson process, can be found in [110, 163].
Markov additive processes are typically implemented in the fields of risk manage-
ment, financial, queueing theory and reliability. [150] considered a Markov additive pro-
cess in the risk and queueing applications and studied the hitting probabilities at upper lev-
els, where the additive process linearly increases or decreases when the background state
is unchanged, and the process may have upward jumps at the transition instants. Asymp-
totic behavior of the ruin probability was studied when the initial reserve goes to infinity,
given that the distributions of claim sizes have light tails. [63] studied the extremes of a
continuous time Markov additive process with one-sided jumps. The Laplace transforms
of extreme distributions were given in terms of two matrices. The results were applied in
determining the steady-state buffer-content distribution of several single-station queueing
systems. However, the underlying work such as solving a nonlinear matrix equation is
intractable.
In degradation analysis, [40] first introduced Markov additive processes in modeling
a wear process and shocks under a Markov dynamic environment. In the model, the shock
arrival rate, the shock sizes and the wear out rate were governed by a random environment,
which evolves as a Markov process; and the Levy´ processes were proposed to represent the additive component. There are no explicit results for reliability characteristics due to the complexity of the evolution of Markov additive processes. [80] analyzed the optimal replacement policy by considering a Markov additive process to model the shocks that occur continuously during a time interval.
29 Recently, special cases of Markov additive processes, Markov-modulated linear and compound Poisson processes were studied in degradation analysis. Explicit results of reliability characteristics were derived by [112–115]. Considering the same model, [84] calculated the limiting average availability of a system, assuming that such system is main- tained through inspection and perfect repair when it fails.
30 Chapter 3
Life Distribution Analysis Based on Levy´ Subordinators for Degrada- tion with Random Jumps
In this chapter, for a component or a system subject to stochastic degradation with sporadic jumps that occur at random times and have random sizes, we propose to model the cumulative degradation with random jumps using a single stochastic process based on the characteristics of Levy´ subordinators, the class of non-decreasing Levy´ processes. Based on the inverse Fourier transform, we derive a new closed-form reliability function and probability density function for lifetime, represented by Levy´ measures. The reliability function derived using the traditional convolution approach for common stochastic models such as gamma degradation process with random jumps, is revealed to be a special case of our general model. Numerical experiments are used to demonstrate that our model performs well for different applications, when compared with the traditional convolution method. More importantly, it is a general and useful tool for life distribution analysis of stochastic degradation with random jumps in multi-dimensional cases.
31 3.1 Introduction
Engineering systems usually deteriorate and lose their intended functionality due to wear, fatigue, erosion, corrosion and aging. The continuous deteriorating process com- monly experiences sporadic jumps due to discrete damages. Stochastic processes are typically used to represent the inherent statistical uncertainty of a degradation process, e.g., compound Poisson process, gamma process, Wiener process. However, there is a lack of research on using a single stochastic process to describe degradation with random jumps. Degradation with random jumps is a process of stochastically continuous degra- dation with sporadic jumps that occur at random times and have random sizes. In this chapter, we intend to model the overall change volume of degradation with random jumps using one stochastic process based on the characteristics of Levy´ subordinators, the class of non-decreasing Levy´ processes. Based on the inverse Fourier transform, we derive a new closed-form reliability function and probability density function (pdf) for lifetime of a component or a system subject to a degradation process with random jumps. The relia- bility function is constructed and represented by a certain Levy´ measure corresponding to a certain Levy´ degradation process.
For systems subject only to sporadic jump damages, a compound Poisson process, a stochastic process with independent and identically distributed (i.i.d.) jumps that oc- cur according to a Poisson process, is one of the appropriate candidates to model the cumulative damages. For degradation due to wear only, a gamma process and a Wiener process are good candidates to model a wear process. The gamma process is suitable for
32 modeling degradation that progresses in one direction due to its property of independent and nonnegative increments. A Wiener process is appropriate for modeling degradation that changes non-monotonically because it can have non-negative and negative increments alternately.
In practice, however, few systems experience a pure sporadic jump damage pro- cess or degradation only. Due to random covariates, a degradation process is typically impacted by sporadic jump damages. By considering degradation with random jumps, a typical approach to calculate reliability is using convolution formula [158]. However, when the wear process has different probability laws from the random jump damages, the calculation becomes less straightforward. For example, when we use a Wiener process
(or a gamma process) for wear and a compound Poisson process for sporadic jumps with normally-distributed jump sizes (or gamma-distributed jump sizes), it is straightforward to derive the reliability function by using convolution; however, the calculation becomes more complex when we consider a Wiener degradation process with gamma jumps. [197] used a gamma process to model wear and a Poisson process with jump sizes following a peaks-over-threshold distribution to model random loads, and the computation of reli- ability is relatively extensive. In addition, the traditional gamma-Poisson-based models may not be suitable enough to fit the general degradation data, especially when there are complex jump mechanisms that cannot be well described by gamma or Poisson laws.
In order to overcome the aforementioned problems, in this chapter, we propose to model a nondecreasing degradation process with random jumps using a single Levy´ sub- ordinator. In our model, we can specify different Levy´ measures to describe different
33 jump mechanisms in degradation, which enables our methods to fit many different types of degradation data sets. By using the inverse Fourier transform, we further derive the closed-form reliability function and pdf of lifetime for a system or a component subjec- t to a degradation process with random jumps, represented by the Levy´ measure. The calculation for reliability is simple enough to be implemented in practice. More impor- tantly, based on mathematical theories in multi-dimensional Levy´ measures, our model in this chapter provides a new framework to analyze multi-degradation processes in multi- component systems.
The organization of the chapter is as follows. Section 3.2 begins with the key no- tions of the general Levy´ process, and then introduces the special cases of Levy´ processes typically used in the literature. In Section 3.3, we derive the reliability function and pdf of lifetime for systems subject to degradation with random jumps described by Levy´ sub- ordinators, based on the Fourier inversion theorem. Section 3.4 studies the reliability of temporally homogenous gamma degradation with different random jumps, a special case of Levy´ subordinators. Numerical examples are illustrated in Section 3.5, and conclusions are given in Section 3.6.
Notation
• Euclidean space: RK ,K ∈ N
K P • Inner product: hx, yi = xiyi i=1
K 1/2 1/2 P 2 • Euclidean norm: |x| = hx, xi = xi i=1
• Borel probability measures: M RK
34 • Borel σ-algebra: B RK
• Levy´ process: X˜ (t)
˜ • Levy´ subordinator: Xs (t)
• Characteristic function: φ
• Levy´ measure: ν
• Levy´ symbol: η
• Standard Brownian motion or Wiener process: B (t)
• Temporally homogeneous gamma process: G (t)
• Compound Poisson process: C (t)
• Indicator function: Ix∈S
• Convolution of finite measures: µ1 ∗ µ2
3.2 Preliminaries of Levy´ Processes
Levy´ processes are stochastic processes whose increments in nonoverlapping time intervals are independent and stationary in time. Their importance in modelling degrada- tion processes stems from [17, 168]: 1) they are analogues of random walks in continuous time; 2) they form special subclasses of Markov processes, for which the analysis is much simpler and provides a valuable guidance for the general case; 3) they are the simplest
35 examples of random motion whose sample paths are right-continuous and have a number
(at most countable) of random jump discontinuities occurring at random times, on each
finite time interval; and 4) they include a number of important processes as special cases,
such as Wiener, compound Poisson, negative binomial, gamma, inverse Gaussian, recipro-
cal gamma, positive hyperbolic, positive hyperbola, positive stable, and positive tempered
stable processes, etc. Therefore, Levy´ process can serve as an important tool for the study
of degradation-based reliability theory. In this section, we introduce Levy´ processes along
with their properties and characteristics on Euclidean space, where the increments can be
positive or negative.
3.2.1 Characteristics
To make our model general, and provide a framework for multi-degradation process-
es, we introduce Levy´ processes on Euclidean space. Definition 3.1 ([17]). X˜ (t) , t ≥ 0 is a Levy´ process defined on a probability space
(Ω, F,P ) , Ω ∈ RK , F ∈ B RK , P ∈ M RK , if
• X˜ (0) = 0 with probability of 1;
˜ • X (t) has independent and stationary increments: for n ∈ N and 0 ≤ t1 < t2 <
˜ ˜ ··· tn+1 < ∞ , the random variables X (ti+1) − X (ti) , 1 ≤ i ≤ n are indepen-
dent and the distribution of X˜ (s + t) − X˜ (s) does not depend on s;
• X˜ (t) is stochastically continuous: for all ε > 0, s > 0,
lim P |X˜ (t) − X˜ (s) | > ε = 0. t→s
36 Characteristic functions are a primary tool for analysis when the distributions have
no analytic forms, especially for Levy´ processes. On Euclidean space, let φX (u) =
R ihu,xi ihu,Xi RK e PX (dx) = E e denote the characteristic function of a random variable √ K X, where PX is the distribution function of X, i = −1 and u ∈ R . More generally, if
µ ∈ M RK , the set of all Borel probability measures on RK , then
Z ihu,xi φµ (u) = e µ (dx) . RK
Characteristic functions have many useful properties, and readers can refer to [137] for more details. One important property that can be used to analyze the sum of indepen- dent variables is described in Definition 3.2.
K Definition 3.2 ([17]). The convolution µ of two finite measures µ1 and µ2 on R , denoted by µ = µ1 ∗ µ2, is a measure defined by
ZZ K µ (B) = Ix+y∈Bµ1 (dx) µ2 (dy) ,B ∈ B R , RK ×RK where B RK is the Borel σ-algebra on Euclidean space.
If X ∼ µ1,Y ∼ µ2, and X and Y are independent, then X + Y ∼ µ , and
φX+Y (u) = φX (u) φY (u) , which implies that the characteristic function of the sum of independent random variables is the product of the characteristic functions of individual random variables. Characteristic functions of Levy´ processes are characterized by Levy´ measures or Levy´ symbols. Next we give the definition of Levy´ measure, and a Levy´ symbol can be represented by a Levy´ measure.
37 Definition 3.3 ([17]). A measure ν on RK is a Levy´ measure if Z min{1, |x|2}ν (dx) < ∞, ν ({0}) = 0. RK Based on the Levy´ Khintchine formula, Levy´ process X˜ (t) has a specific form for
its characteristic function. More precisely, for all t ≥ 0, x ∈ RK ,
ihu,X˜(t)i tη(u) φX˜(t) (u) = E e = e , (3.1)
where Z 1 ihu,xi η (u) = ihb, ui − hu, aui + e − 1 − I|x|<1ihu, xi ν (dx) 2 RK is the Levy´ symbol, in which ν is the Levy´ measure, b is a constant on RK , and a is a positive definite symmetric K ×K matrix. Levy´ measure is the most important element of a Levy´ process, in a sense that if we specify a Levy´ measure, we can get the corresponding
Levy´ process and its characteristic function.
Lemma 3.1 ([168]). Let K=1. A Levy´ process is a subordinator if and only if a = 0,
R ν (dx) = 0, R min{1, x}ν (dx) < ∞, and the drift b − R xν (dx) ≥ 0. (−∞,0) R+ 0 Levy´ subordinators form the class of nondecreasing Levy´ processes, taking values in R+ ≡ (0, ∞). Based on (3.1) and Lemma 3.1, a one-dimensional Levy´ subordinator ˜ Xs (t) has the characteristic function ˜ φ (u) = E eiuXs(t) = etηs(u), X˜s(t) where Z ∗ iux ηs (u) = ib u + e − 1 ν (dx) , R+ ∗ R is the Levy´ symbol, and b = b − 0 38 3.2.2 Special Cases of Levy´ Processes Levy´ processes are stochastic processes with independent and stationary increments over time. Some special Levy´ processes have been widely used to model degradation processes in the literature, such as Wiener process and gamma process for wear, and com- pound Poisson process for pure jump damages. The Levy´ measures and Levy´ symbols for these common special cases are introduced in this section. 3.2.2.1 Linear Process When a = ν = 0, b 6= 0, the Levy´ symbol in (3.1) becomes η (u) = ihb, ui, and the ithb,ui ˜ characteristic function in (3.1) is φX˜(t) (u) = e , indicating that X (t) = bt, where b is a constant. Therefore, X˜ (t) is a deterministic linear process, which is not suitable for modeling stochastic degradation processes. 3.2.2.2 Brownian Motion/Wiener Process 1 When a 6= 0, b 6= 0, ν = 0, the Levy´ symbol becomes η (u) = ihb, ui − 2 hu, aui 1 t[ihb,ui− 2 hu,aui] and φX˜(t) (u) = e , which is the characteristic function of Brownian motion with drift b. The case a = 1, b = 0, ν = 0 is usually called standard Brownian motion or Wiener process (B (t) , t ≥ 0) , which has a Gaussian density 1 |x|2 − 2t ρt (x) = K e . (2πt) 2 Wiener process and Brownian motion with drift are not suitable for modeling monotoni- cally increasing wear processes, because their increments are not always positive. 39 3.2.2.3 Temporally Homogeneous Gamma Process When a = 0, b 6= 0, ν 6= 0, if ν is a finite measure, we have Z η (u) = ihb∗, ui + eihu,xi − 1ν (dx) , RK ∗ R where b = b − 0<|x|<1 xν (dx). Next, we find the special form for ν to obtain the temporally homogeneous gamma process. For a gamma process on R ≡ R1, if the shape parameter α (t) = αt, t ≥ 0 (i.e., the second condition in Definition 3.1 is satisfied), it is a temporally homogeneous gamma βαtxαt−1e−βx process G(t). G (t) has a density fG(t) = Ga (x|αt, β) = Γ(αt) , x > 0, t ≥ 0. Then the characteristic function of G (t) can be expressed as αt Z ∞ −βx β β iux e φG(t) (u) = = exp αt ln = exp αt e − 1 dx . β − iu β − iu 0 x Therefore, the temporally homogeneous gamma process is a special case of Levy´ process, and its Levy´ measure is ν(dx) = αx−1e−βxdx, and Levy´ symbol is η (u) = R ∞ iux e−βx ∗ α 0 (e − 1) x dx, with a = 0, b = 0. The temporally homogeneous gamma pro- cess is a Levy´ process that is always positive and strictly increasing, and it is suitable for modeling strictly increasing wear processes with a linear mean path, αt/β. 3.2.2.4 Compound Poisson Process For a Poisson process with parameter λ, N (t) ∼ Poisson(λt), and P (N(t) = n) = n e−λt(λt) ∗ n! , for n = 0, 1, 2, ··· . Let (J (n) , n ∈ N(t)) be the jump size described by a sequence of i.i.d. random variables taking values in R with distribution µJ , which is 40 independent of N(t). The compound Poisson process C (t) is defined as follows: C (t) = J ∗ (1) + ··· + J ∗ (N (t)) . Based on Definition 3.2, we obtain the characteristic function of compound Poisson process: ∞ n ! ∞ n iu P J∗(k) −λt iuC(t) X X e (λt) n φ (u) = E e = P (N (t) = n)E e k=1 = φ (u) C(t) n! J n=0 n=0 Z iux = exp (λt (φJ (u) − 1)) = exp λt e − 1 µJ (dx) . R Therefore, for a compound Poisson process C (t), the Levy´ measure is ν(dx) = λµJ (dx), and the Levy´ symbol is Z iux ηC (u) = e − 1 λµJ (dx) . R The sample paths of C (t) are piecewise constant on finite intervals with jump dis- continuities at random times. It is suitable for modeling pure jump damages. 3.3 Life Distribution Analysis Based on Levy´ Subordinators ˜ We use the Levy´ subordinator Xs(t) to represent the monotonically nondecreasing volume of degradation with random jumps up to time t. A component or a system fails ˜ when Xs(t) exceeds a failure threshold x, assuming that it subjects to one degradation ˜ process that begins with Xs(0) = 0. To simplify the formula, we assume the failure threshold is a constant, and it is straightforward to extend the model when the failure threshold is a random variable. 41 The lifetime of the system is defined as the first passage time of the stochastic pro- cess: ˜ Tx = inf{t : Xs(t) > x}. ˜ Since Xs(t) is nondecreasing, we have ˜ {Tx ≥ t} ≡ {Xs(t) ≤ x}. Then the reliability function can be defined as R (t) = P (T ≥ t) = P X˜ (t) ≤ x = F (x) . x s X˜s(t) (3.2) The relationship in (3.2) holds for a broad class of stochastic processes that have monotonically nondecreasing paths. In this section, we present a method based on the inverse Fourier transform to derive the reliability function for systems subject to a degra- dation process with jumps that can be described by a Levy´ subordinator. Although the pdf of Levy´ process is not readily available, we have the expression of its characteristic function. Since there is a one-to-one correspondence between the cumulative distribution function (cdf) and the characteristic function, we can obtain one of them if the other one is known. Based on the Fourier inversion theorem, Shephard [173] provided the following remarkable theorems describing the cdf as the function of φ(u) for a random variable. Lemma 3.2 ([173]). If the probability density function f and the characteristic function φX (u) are integrable in the Lebesgue sense, then under the assumption that the mean for the random variable of interest exists, the following equality holds: 1 1 Z ∞ e−iux FX (x) = − ∆ φX (u) du, 2 2π 0 u iu 42 where ∆ ρ (u) = ρ (u) + ρ (−u). u The following Lemma 3.3 is the multivariate generalization of Lemma 3.2. Lemma 3.3 ([173]). If the probability density function f and the characteristic function φX (u) are integrable in the Lebesgue sense, then under the assumption that the mean for the multi-dimensional random variable of interest exist, the following equality holds: K Z ∞ Z ∞ −ihu,xi (−2) e ∗ K ··· ∆ ∆ ··· ∆ φX (u) du = z (x) , (2π) 0 0 u1 u2 uK iu1iu2 ··· iuK where ∗ K K−1 z (x) = 2 F (x1, ··· xK ) − 2 (F (x2, x3, ··· , xK ) + ··· + F (x1, ··· , xK−2, xK−1)) K−2 K + 2 (F (x3, x4, ··· , xK ) + ··· + F (x1, ··· , xK−3, xK−2)) + ··· + (−1) . The mean’s existence is a sufficient but not a necessary condition for the results to hold. It can be removed by using the principal value of the integral [87]. Lemma 3.2 turns out to be a special case of Lemma 3.3 that deals with multi-dimensional variables. For an example of two-dimensional variables, if we know the characteristic function φX (u), ∗ we can get the expression of z (x1, x2) = 4F (x1, x2) − 2F (x1) − 2F (x2) + 1 based on Lemma 3.3. If we know the characteristic function of each variable, φX1 (u1) and φX2 (u2), we can have F (x1) and F (x2) based on Lemma 3.2. Finally, we can solve for the joint distribution function F (x1, x2) of X1 and X2. Integration rules for the computation of the multivariate distribution function are described in [174]. In this chapter, we focus on one-dimensional Levy´ degradation processes. When ˜ K = 1, for all t ≥ 0, u ∈ R, the characteristic function of a Levy´ subordinator Xs(t) is 43 ˜ available. For Xs(t), we derive the reliability function and pdf of lifetime in the following corollaries, respectively. Corollary 3.1. For systems subject to stochastic degradation with random jumps that can be described by Levy´ subordinators, assuming the failure threshold value is x, the reliability function represented by the Levy´ measure is 1 1 Z ∞ e−iux Z R (t) = − ∆ exp t ib∗u + eiux − 1ν (dx) du, (3.3) u 2 2π 0 iu R+ ∗ R where b = b − 0 Corollary 3.2. For systems subject to stochastic degradation with random jumps that can be described by Levy´ subordinators, assuming the failure threshold value is x, the pdf of lifetime represented by the Levy´ measure is ∂R (t) f (t) = − ∂t −iux ∗ R iux Z ∞ e exp t ib u + (e − 1)ν (dx) (3.4) 1 R+ = ∆ −1 du, 2π u 0 iu ib∗u + R (eiux − 1)ν (dx) R+ ∗ R where b = b − 0 For systems subject to degradation with random jumps that can be described by Levy´ subordinators, we can first specify a certain Levy´ measure and then calculate the reliability function and pdf using (3.3) and (3.4). [112] gave explicit results for wear processes in Markovian environment, which requires to use multi-inverse algorithms to calculate. Al- though they are not explicit, our results in (3.3) and (3.4) can be computed comparatively cheap based on [173]. The advantages of our results are twofold: 1) they are general 44 because we can specify different Levy´ measures to fit different types of degradation data sets, while the models in the literature become special cases of our models, and 2) they provide a methodology to deal with complex random jumps in degradation processes, e.g., when the distributions of jumps size are not additive. 3.4 Life Distribution Analysis for Temporally Homogeneous Gamma Process with Random Jumps To demonstrate the advantages of our models, we present the life distribution analysis for a degradation process represented by a sum of a temporally homogeneous gamma pro- cess and a compound Poisson process. For this case, we can get a closed-form reliability function using the traditional convolution approach, which enables the direct comparison with our proposed method. Nevertheless, we can also choose different Levy´ measures to illustrate our general model, such as a Levy´ measure for a sum of a positive tempered stable process and a compound Poisson process. In this case, however, no comparison can be directly made, because the traditional convolution approach cannot generate a closed- form reliability function considering that the positive tempered stable process does not have a closed-form distribution. 45 3.4.1 Reliability Function Using Traditional Convolution Approach We first present the reliability function derived from the traditional convolution ap- proach for the temporally homogeneous gamma process with random jumps. If a degra- dation process with sporadic jumps can be well described by the sum of a temporally homogeneous gamma process and a compound Poisson process that are independent, the traditional convolution approach provides the reliability function: N(t) ˜ X ∗ R (t) = P Xs (t) ≤ x = P G (t) + Ji ≤ x i=0 ∞ N(t) X X ∗ (3.5) = P G (t) + Ji ≤ x |N (t) = n P (N (t) = n) n=0 i=0 ∞ n ! n X X e−λt(λt) = P G (t) + J ∗ ≤ x . i n! n=0 i=0 n ∗ ∗ ∗ P ∗ If the jump size follows a gamma distribution, Ji ∼ Gamma(α , β ), then Ji ∼ i=0 Gamma(nα∗, β∗). If the scale parameter of G (t) is the same as β∗, i.e., β = β∗, then n P ∗ ∗ G (t) + Ji ∼ Gamma(αt + nα , β). The reliability function for this special case is i=0 ∞ n X Γ(αt + nα∗, xβ)e−λt(λt) R (t) = 1 − , (3.6) Γ(αt + nα∗) n! n=0 ∗ R ∞ αt+nα∗−1 −y ∗ R ∞ αt+nα∗−1 −y where Γ(αt + nα ) = y=0 y e dy , Γ(αt + nα , xβ) = y=xβ y e dy, and x is the threshold value. We can see that (3.6) is derived based on two assumptions: 1) the jump size follows ∗ a gamma distribution, and 2) the scale parameters of G(t) and Ji are the same. When β 6= β∗, it is still manageable to derive the reliability function using the convolution of 46 ∗ gamma distributions [106]. However, if Ji follows a different distribution than a gamma distribution (such as inverse Gaussian distribution, Levy´ distribution, Pareto distribution, or stable distribution), it becomes complex to calculate the reliability function in (3.5). Our approach in Corollary 3.1 is capable of dealing with these cases by using the Levy´ measure. 3.4.2 Reliability Function Using Levy´ Measures In this section, we use our new approaches in Corollaries 3.1 and 3.2 to derive the reliability function and pdf for a temporally homogeneous gamma process with random jumps. As given in Section 3.2.2.3, Levy´ measure for a temporally homogeneous gamma −1 −βx process is ν1(dx) = αx e dx, and Levy´ measure for a compound Poisson process is ν2(dx) = λµJ (dx). Then for a Levy´ subordinator that is sum of a temporally homo- geneous gamma process and a compound Poisson process, the characteristic function is derived to be Z αe−βx φ (u) = φ (u) = exp t eiux − 1 + λµ0 dx , X˜s(t) G(t)+C(t) J R+ x 0 where µJ is the pdf of the jump size. Therefore, we can model this Levy´ subordinator −1 −βx by specifying its Levy´ measure to be ν = ν1 + ν2 = αx e dx + λµJ (dx). Based on Corollary 3.1, the reliability function is Z ∞ −iux Z −βx 1 e iux αe 0 R (t) = − ∆ exp t e − 1 + λµJ dx du. u 2 0 2πiu R+ x (3.7) 47 Based on Corollary 3.2, the pdf of lifetime is ∂R (t) f (t) = − ∂t αe−βx ∞ −iux R iux 0 (3.8) Z e exp t R (e − 1) x + λµJ dx = + du. ∆ −1 0 u 2πiu R (eiux − 1) αe−βx + λµ0 dx R+ x J The results in (3.7) and (3.8) can be applied to the jump size following a general distribution µJ defined on [0, ∞). In the following, we derive the reliability function and pdf for three different jump types. 3.4.2.1 Gamma-distributed Jump Sizes If the jump size follows a gamma distribution, βα∗ xα∗−1e−β∗x µ0 = Ga (x|α∗, β∗) = , x > 0, J Γ(α∗) then the characteristic function for the compound Poisson process is Z α∗ α∗−1 −β∗x iux β x e φC(t) (u) = exp λt e − 1 ∗ dx R+ Γ(α ) Z α∗ α∗−1 −β∗x Z α∗ α∗−1 −β∗x iux β x e β x e = exp λt e ∗ dx − ∗ dx R+ Γ(α ) R+ Γ(α ) Z α∗ α∗−1 −β∗x iux β x e = exp λt e ∗ dx − 1 R+ Γ(α ) ∗ !! β∗ α = exp λt − 1 . β∗ − iu Then the reliability function in (3.7) is ∗ !!! 1 Z ∞ e−iux β αt β∗ α R (t) = − ∆ exp λt ∗ − 1 du. (3.9) 2 0 u 2πiu β − iu β − iu (3.9) is a general formula for reliability function of systems subject to degradation described by the sum of a gamma process and a compound Poisson process with gamma- distributed jumps, regardless of β = β∗ or not, while (3.6) is only valid for the case of 48 β = β∗. The pdf of lifetime in (3.8) is derived to be ∂R (t) f (t) = − ∂t ∗ ∗ α αt λt β −1 −iux β β∗−iu Z ∞ e β−iu e = ∆ du. α α∗ −1 0 u β β∗ 2πiu ln β−iu + λ β∗−iu − 1 3.4.2.2 Levy-distributed´ Jump Sizes If the jump size follows a different distribution than a gamma distribution, we can also derive the reliability function and pdf using (3.7) and (3.8). When the jump size follows a Levy´ distribution, the pdf is given as [16] q ξ exp(− ξ ) 2π 2(x−$) 3 for x > $ > 0 0 2 µJ (x; $, ξ) = (x − $) 0 otherwise. Levy´ distribution is a continuous probability distribution of a non-negative random variable when $ > 0. It has different probability laws from a gamma distribution, leading to complex calculation in the convolution approach. Since the characteristic function of √ Levy´ distributed variable is eiu$− −2iuξ [92], the reliability function of a gamma process with additional Levy-distributed´ jumps is derived from (3.7) to be ! 1 Z ∞ e−iux β αt √ R (t) = − ∆ exp λt eiu$− −2iuξ − 1 du. 2 0 u 2πiu β − iu The pdf of lifetime in (3.8) is ∂R (t) f (t) = − ∂t αt √ β λt eiu$− −2iuξ−1 ∞ −iux ( ) Z e β−iu e = ∆ du. α √ −1 0 u β iu$− −2iuξ 2πiu ln β−iu + λ e − 1 49 3.4.2.3 Inverse Gaussian-distributed Jump Sizes The pdf of the inverse Gaussian distribution is r 2 ϑ ϑ(x − ς) exp{− } for x > 0 0 2πx3 2ς2x µJ (x; ς, ϑ) = 0 otherwise, where ς > 0 is the mean, and ϑ > 0 is the shape parameter. It has different probability laws from a gamma distribution. Since the characteristic function of an inverse Gaussian- q 2 ϑ 1− 1− 2iuς distributed variable is e ς ϑ , the reliability function of a gamma process with additional inverse Gaussian-distributed jumps in (3.7) is ∞ αt q 2 !!! 1 Z e−iux β ϑ 1− 1− 2iuς R (t) = − ∆ exp λt e ς ϑ − 1 du. 2 0 u 2πiu β − iu The pdf of lifetime in (3.8) is ∂R (t) f (t) = − ∂t √ 2 ϑ 1− 1− 2iuς ς ϑ αt λte −1 β ∞ −iux Z e β−iu e = ∆ −1 du. u q 2 !! 0 α ϑ 1− 1− 2iuς β ς ϑ 2πiu ln β−iu + λ e − 1 Beyond Levy´ measures covered in this section, we can specify additional Levy´ mea- sures for model construction in order to fit the corresponding degradation data. Some inter- −2κ −κ−1 1 2 δγ κx exp(− 2 γ x) esting Levy´ measures have been studied in [25], such as ν(dx) = Γ(κ)Γ(1−κ) dx, x, δ > 0, 0 < κ < 1, γ ≥ 0 for the positive tempered stable process PTS(κ, δ, γ). 50 3.5 Numerical Examples We consider the crack growth process in a device, which is subject to degradation due to fatigue and a variety of overloads that can occur in manufacturing, deployment, and ˜ operation phases. We use a Levy´ subordinator Xs(t) to represent the growth of a crack at time t, specifically, a temporally homogeneous gamma process with random jumps. −1 −βx Then the Levy´ measure is ν = αx e dx + λµJ (dx). In particular, we consider three different distributions to model the jump size: gamma, Levy´ and inverse Gaussian. The specific values for the parameters are given in Table 3.1. A device fails when the crack length exceeds the threshold x. Table 3.1: Parameter values for models in Chapter 3 Parameters Values Parameters Values α 5 $ 1 β 0.8 ξ 0.002 λ 3 ς 1 α∗ 1 or 10 ϑ 1 β∗ 0.8 or 15 x 50 51 1 0.9 0.8 0.7 0.6 0.5 Reliability 0.4 0.3 0.2 Traditional method:beta*=beta=0.8 0.1 New method:beta*=beta=0.8 New method:beta*=15 0 0 1 2 3 4 5 6 7 8 9 10 t Figure 3.1: Reliability function for gamma degradation with additional gamma jumps 1 0.9 0.8 0.7 0.6 0.5 Reliability 0.4 0.3 0.2 Gamma-type 0.1 Levy-type Inverse Gaussian-type 0 0 1 2 3 4 5 6 7 8 9 10 t Figure 3.2: Reliability function for gamma degradation with three jump types 52 0.5 Gamma-type 0.45 Levy-type Inverse Gaussian-type 0.4 0.35 0.3 0.25 Pdf 0.2 0.15 0.1 0.05 0 0 1 2 3 4 5 6 7 8 9 10 t Figure 3.3: Pdf of lifetime for gamma degradation with three jump types 5 Gamma-type 4.5 Levy-type Inverse Gaussian-type 4 3.5 3 2.5 Hazard rate 2 1.5 1 0.5 0 0 1 2 3 4 5 6 7 8 9 10 t Figure 3.4: Hazard rate for gamma degradation with three jump types 53 Figure 3.1 shows the reliability function over time of devices subject to a gamma degradation with additional gamma-distributed jumps. When the parameter β∗ = β = 0.8, both the traditional convolution approach in (3.6) and our proposed model in (3.9) can solve the problem, showing the same curve of R(t). When β∗ 6= β, the reliability curve is provided by our model in (3.9), and the convolution approach becomes complex in this case. Figure 3.2 shows the reliability functions over time of devices subject to a gamma degradation with three different jump distributions. It demonstrates that for Levy´ and inverse Gaussian distributed jump sizes, which have different probability laws from gam- ma, we can readily calculate the reliability by using our proposed model. Figure 3.3 and Figure 3.4 illustrate the pdfs and hazard rates for the lifetime of the devices. As expected, we can see that the hazard rates increase over time for all three cases with nondecreasing degradation processes. 3.6 Conclusions One of the challenging aspects in reliability analysis is how to formulate the reli- ability function from a degradation process that a system or a component experiences. In this chapter, we presented a novel model concerning the stochastic mechanism of a complex degradation process that also subjects to random jumps. Based on inverse Fourier transforms, the reliability function and pdf of lifetime were derived. Our model is general because we can specify different Levy´ measures to fit different types of degradation data 54 sets, and the models in the literature become special cases of our model. In addition, by providing a methodology to deal with complex random jumps in degradation processes, our method can solve the problems that the traditional convolution method cannot readily solve, i.e., when the distribution of jumps sizes is not additive. Our new method provides a convenient and general way to evaluate the system reliability. The analysis for reliability is simple enough to be implemented in practice. More importantly, the model provides a framework for reliability analysis of multi- degradation processes in multi-component systems. To derive the reliability function for multi-Levy´ degradation processes on RK , we need to construct multi-dimensional Levy´ measures. If the multi-degradation processes are dependent, the construction of the multi- dimensional Levy´ measures can refer to Levy´ copula theory [55, 109], which will be explored in Chapter 6. In order to apply the model to degradation data analysis, statistical inference on Levy´ measures is another research direction, which will be developed in Chapter 7. Levy´ subordinators studied in this chapter are a class of homogeneous Markov pro- cesses with a linear mean path over time t. For non-linear degradation paths, special non- homogeneous subordinators (e.g., non-homogeneous gamma processes) have been studied in the literatures [154, 201], which are essentially non-decreasing additive processes. Our model can be readily extended to the class of non-homogeneous subordinators when the Levy´ measures are available. 55 Chapter 4 Levy´ Driven Non-Gaussian Ornstein-Uhlenbeck Processes for Degrada- tion-based Reliability Analysis In this chapter, we further explore Levy´ subordinators and extend the model using Levy´ driven non-Gaussian OU processes. To obtain explicit results of reliability func- tion and lifetime moments, we propose to use Fokker-Planck equations for both Levy´ subordinators and their corresponding OU processes. The most important advantage of the models stems from the flexibility of such processes in modeling stylized features of degradation data series such as jumps, linearity/nonlinearity, symmetry/asymmetry, and light/heavy tails. Numerical experiments are used to demonstrate that our general models perform well and are applicable for analyzing a great deal of degradation phenomena. More importantly, they provide us a new methodology to deal with degradation processes under dynamic environments. 4.1 Introduction Chapter 3 gave a new closed-form of reliability function for degradation described by Levy´ subordinators, a class of non-decreasing Levy´ processes, which is consistent with some observed physical degradation phenomena. The advantages of using Levy´ subor- dinators were also demonstrated. With independent and stationary increments, however, 56 all Levy´ processes have linear mean paths, i.e., the mean of Levy´ processes is linear with respect to (w.r.t) time. To overcome the limitation from the linear mean property, Gaussian OU processes driven by a Wiener process have been developed for survival analysis [1]. However, the assumptions of no jumps and Gaussian distribution (symmetric and light- tailed, i.e., all the positive moments are finite) are not consistent with many degradation phenomena. In this chapter, to flexibly handle stylized features of degradation data series such as complex jumps, linearity/nonlinearity, symmetry/asymmetry, and light/heavy tails, we propose to model stochastic degradation with independent or dependent increments us- ing Levy´ subordinators or OU processes driven by Levy´ subordinators (i.e., non-Gaussian OU processes), respectively. To the best of our knowledge, non-Gaussian OU processes have not been used in degradation modeling. In fact, it is nontrivial to obtain a closed- form distribution function for an OU process driven by a Levy´ process. For these general stochastic degradation processes, we construct systematic procedures to derive the explicit expressions for reliability function and lifetime moments using Fokker-Planck equations. Our proposed new models offer a general approach for modelling stochastic degradation with complex jump mechanisms using a broad class of Levy´ processes and their functional extensions. Fokker-Planck equations provide us a way to analyze probability laws for stochastic processes, especially for those without closed-form distributions. Fokker-Planck equa- tions represent the fascinated research work from mathematicians in the field of stochastic processes. As the partial differential equations (PDE) of the probability density function- s, they describe the time evolution of probability density for stochastic processes, and 57 are thus useful in quantifying random phenomena, such as propagation of uncertainty. We can find the Fokker-Planck equations for Weiner-based processes in many textbooks [117, 165]. For such processes, it is straightforward to derive the Fokker-Planck equations, because of the absence of jump mechanisms. However, for Levy-based´ processes, explicit results of Fokker-Planck equations cannot be easily derived, due to the difficulty in ob- taining the expression for the adjoint operators of the infinitesimal generators associated with Levy-based´ processes [187]. Some interesting results of Fokker-Planck equations for Levy-based´ processes are in [61, 169, 187], and [164] gave a numerical algorithm to calculate the mean exit time for Levy´ systems. In this chapter, we consider a single degradation process with random jumps in a system, i.e., a process of stochastically continuous degradation with sporadic jumps that occur at random times and have random sizes. The system fails when the degradation process hits a boundary. We first use Levy´ subordinators, a class of Levy´ processes with non-decreasing sample paths, to model the evolution of degradation with linear mean paths (Figure 4.1: Wiener process (0, 1); gamma process (20, 20); compound Poisson process with jump density 2 and jump size following gamma distribution (1, 10); Levy´ subordinator a: inverse Gaussian process (0.5, 0.1); Levy´ subordinator b: positive stable process (0.9); Levy´ subordinator c: positive stable process (0.92)). We then propose a functional extension of Levy´ subordinators, non-Gaussian OU processes (OU processes driven by Levy´ subordinators), to model degradation processes with nonlinear mean paths (Figure 4.2: non-Gaussian OU process a: OU process driven by inverse Gaussian process (0.5, 0.1) and α = 0.2; non-Gaussian OU process b: OU process driven by positive stable 58 process (0.9) and α = 0.1; non-Gaussian OU process c: OU process driven by positive stable process (0.92) and α = 0.1). Wiener process Gamma process Compound Poisson process X(t) X(t) X(t) −0.5 0.0 0.5 1.0 1.5 2.0 2.5 0 2 4 6 8 10 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 t t t Levy subordinator a Levy subordinator b Levy subordinator c X(t) X(t) X(t) 0 5 10 15 0 2 4 6 8 10 12 14 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 t t t Figure 4.1: Sample paths of Levy´ processes Figure 4.1 shows sample paths of three commonly-used Levy´ processes (Wiener pro- cess, gamma process, and compound Poisson process), and three Levy´ subordinators with different jump mechanisms specified by different Levy´ measures. Figure 4.2 illustrates sample paths of OU processes driven by Levy´ subordinators, a class of non-Gaussian OU processes, and they are the solutions of stochastic differential equations (SDE) driven by Levy´ subordinators. The sample data are simulated using R(YUIMA) [33]. In practice, many degradation processes in highly reliable systems have similar paths to those in Figure 4.2: they increase slowly at the early stage, but increase sharply when the degradation is 59 accumulated. In these cases, the linear mean path of a Levy´ subordinator is not appropriate to represent the degradation. Non−Gaussian OU a Non−Gaussian OU b Non−Gaussian OU c Y(t) Y(t) Y(t) 0 10 20 30 40 0 5 10 15 20 25 30 35 0 10 20 30 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 t t t Figure 4.2: Sample paths of non-Gaussian OU processes For both general Levy´ subordinators and non-Gaussian OU processes, the proba- bility distributions are not analytically available. In addition, the analytical derivation is intractable for non-Gaussian OU processes. In this chapter, we tackle these challenges by using the corresponding Fokker-Planck equations, and then derive explicit expressions for reliability function and lifetime moments in terms of Laplace transform. The results are compact enough to compute and evaluate reliability characteristics conveniently. More importantly, by introducing Fokker-Planck equations to stochastic degradation analysis, our work provides a new methodology for reliability analysis of complex degradation phenomenon, such as multi-degradation processes under dynamic environments. The organization of this chapter is as follows. Section 4.2 begins with the famous Levy-It´ oˆ decomposition, and then describes the model construction. In Section 4.3, we derive the explicit expressions of reliability function and lifetime moments for systems subject to degradation described by Levy´ subordinators and non-Gaussian OU processes, 60 respectively, based on Fokker-Planck equations. Numerical examples are illustrated in Section 4.4, and conclusions are given in Section 4.5. 4.2 Preliminaries 4.2.1 Levy-It´ oˆ Decomposition We begin with the definition of Poisson random measure from [168]. A random variable J has a Poisson distribution with a mean 0 if J = 0 almost surely (a.s.); and J has a Poisson distribution with a mean +∞ if J = +∞ a.s.. ¯ Definition 4.1 ([168]). Let (Λ, B, ν) be a σ-finite measure space. Given Z+ = ¯ {0, 1, 2, ··· , +∞}, a family of Z+-valued random variables {J (A): A ∈ B} is called a Poisson random measure on Λ with an intensity measure ν, if the following conditions hold: • For every A, J(A) has a Poisson distribution with a mean ν(A); • If A1,A2, ··· An are disjoint, then J(A1),J(A2), ··· J(An) are independent; and • For every w, J(·, w) is a measure on Λ. Lemma 4.1 (The Levy-It´ oˆ Decomposition [17]). If X˜(t) is a Levy´ process, then there exist K b ∈ R , a Brownian motion Ba with a covariance matrix a, and an independent Poisson random measure J on R+ × RK such that, for each t ≥ 0, Z Z ˜ X(t) = bt + Ba(t) + y J (t, dy) − ν (t, dy) + yJ (t, dy), |y|<1 |y|≥1 61 where ν (t, dy) is the mean of the Poisson random measure J (t, dy). The intensity measure ν (t, dy) is often called the Levy´ measure. According to the property of independent and stationary increments of Levy´ process, and from Levy-´ Khintchine formula [168], ν (t, dy) = ν (dy) t. 4.2.2 Model Construction We assume there is a single degradation path with random jumps occurring in a ˜ system. We use a Levy´ subordinator Xs (t) and a non-Gaussian OU process Y (t) to model the degradation evolution with linear and non-linear mean paths, respectively. ˜ Based on Lemmas 3.1 and 4.1, for Levy´ subordinator Xs (t), Z Z ˜ Xs(t) = bt + y J (t, dy) − ν (t, dy) + yJ (t, dy), (4.1) 0 R where the continuous degradation is modeled by b − 0 We can specify different forms of Levy´ measures in order to model different complex jump mechanisms. If we specify ν(dx) = αx−1e−βxdx for small jumps in an infinitesimal time interval, then the Levy´ subordinator in (4.1) is a temporally homogeneous gamma process (a gamma process with stationary increments) G (t), which has a density fG(t) = βαtxαt−1e−βx Ga (x|αt, β) = Γ(αt) , x > 0, t > 0. G (t) has an infinite number of small jumps in a finite time interval, thus having infinite activity. The inverse Gaussian process has the similar feature, but with more very small jumps than G (t), since its Levy´ density ν0(x) = ν(dx)/dx approaches to infinity faster than the gamma process does as x goes 62 to zero. Another case is the positive stable process whose probability distribution is in general unknown in simple form. For big jumps occurring according to the Poisson law, we can specify ν(dx) = λµJ (dx), and then the Levy´ subordinator is a compound Poisson process C (t) with a jump density λ and a jump size distribution µJ . C (t) has a finite number of jumps over a finite time interval, i.e., finite activity. Another example of Levy´ subordinators with finite activity is the negative binomial process. Compared to C (t), in the negative binomial process, the interarrival times of jumps are not exponentially distributed and the variance of the number of jumps in a time interval is larger than the corresponding mean [25]. ˜ A non-Gaussian OU process Y (t) is the solution of a SDE driven by Xs(t): ˜ dY (t) = αY (t) dt + dXs (t) . (4.2) Proposition 4.1. The non-Gaussian OU process resulted from (4.2) is Z t αt α(t−ξ) ˜ Y (t) = e Y (0) + e dXs(ξ). 0 Proof. If f(t, y) ∈ C1,2, then based on Taylor series, we have ∂f ∂f 1 ∂2f 1 ∂2f ∂2f df = dt + dy + (dt)2 + (dy)2 + dydt. ∂t ∂y 2 ∂t2 2 ∂y2 ∂y∂t −αt ∂f −αt ∂f −αt ∂2f Let f(t, y) = ye , then ∂t = −αye , ∂y = e , and ∂y2 = 0. We obtain ∂f ∂f df = dt + dy = −αye−αtdt + e−αtdy = e−αtdx, ∂t ∂y and Z t αt αt −αξ yt = e y0 + e e dxξ. 0 63 Y (0) represents the initial state of the degradation, and we assume Y (0) = 0 a.s. as many new systems have not accumulated degradation when they are firstly operated. We assume α > 0, which guarantees that the degradation process is non-decreasing. Y (t) is expressed as Z t α(t−ξ) ˜ Y (t) = e dXs(ξ) 0 Z t Z Z = eα(t−ξ) bdξ + y J(dξ, dy) − ν(dξ, dy) + yJ (dξ, dy) 0 0 b−R yν(dy) 0 the random jumps are modeled by the Poisson random measure R t eα(t−ξ) R yJ (dξ, dy). 0 R+ As illustrated in Figure 4.2, the mean degradation path of Y (t) is exponential w.r.t t, in- ˜ stead of linear of Xs(t). 4.3 Reliability Function and Lifetime Moments ˜ The system fails when the degradation process Xs(t) or Y (t) exceeds a failure thresh- ˜ old x or y. Based on Xs(t), the lifetime and its moments are defined respectively as ˜ n n Tx = inf{t : Xs(t) > x},M(TX , x) = E(Tx ). The reliability function can be defined as R (x, t) = P (T ≥ t) = P X˜ (t) ≤ x = F (x) . X x s X˜s(t) 64 n n Based on Y (t), similar definitions are Ty = inf{t : Y (t) > y},M(TY , y) = E(Ty ), and RY (y, t) = P (Ty ≥ t) = P (Y (t) ≤ y) = FY (t) (y). For many new systems that ˜ have not accumulated degradation when they are firstly operated, we have Xs(0) = 0 a.s., and R (x, 0) = P X˜ (0) ≤ x = F (x) = h(x), X s X˜s(0) ∂F ˜ (x) p(x, 0) = Xs(0) = δ(x), ∂x where h(x) = I[0,∞)(x) is the unit step function (or the Heaviside step function), and δ(x) is the Dirac delta function. Similarly, we have RY (y, 0) = P (Y (0) ≤ y) = FY (0) (y) = h(y), ∂F (y) p(y, 0) = Y (0) = δ(y). ∂y ˜ In addition, RX (0, t) = P (T0 ≥ t) = P Xs (t) ≤ 0 = I(−∞,0](t), and RY (0, t) = P (T0 ≥ t) = P (Y (t) ≤ 0) = I(−∞,0](t). To obtain expressions of reliability functions and lifetime moments, we need to study ˜ the probability laws of Xs(t) and Y (t). Since there are no closed-form distribution func- tions for general Levy´ subordinators, it is a challenge to derive the explicit expressions for reliability functions and lifetime moments. As PDEs of probability density functions, Fokker-Planck equations [187] provide us a way to overcome the challenge in analyzing probability laws for stochastic processes we are interested in, especially for those without closed-form distributions. The Fokker-Planck equation, also known as the Kolmogorov forward equation, describes the time evolution of probability density for stochastic pro- cesses. 65 Let L be an operator, and L∗ be the adjoint operator of L, then Z Z Lf (x)g (x) dx = f (x) L∗g (x) dx. R R Let p(x, t) be the probability density function for a stochastic process X(t), and the Fokker-Planck equation is ∂p(x, t) = L∗p(x, t), ∂t where L∗ is the adjoint operator of the infinitesimal generator L of X(t): E {f (X ) |X = x} − f (x) Lf(x) = lim t+∆t t . ∆t→0 ∆t Laplace transform of p(x, t) w.r.t. t is defined to be Z pL(x, ω) = e−ωtp (x, t) dt, ω > 0. R+ Laplace transform of pL(x, ω) w.r.t. x is Z pLL(u, ω) = e−uxpL (x, ω)dx, u > 0. R+ Lemma 4.2. Let RLL(u, ω) be the Laplace expression of reliability function R(x, t), then RLL(u, ω) = u−1pLL(u, ω). Proof. From the definition of the reliability function, we have ∂R (x, t) p(x, t) = . ∂x The Laplace transform of p(x, t) w.r.t. t is Z Z ∂R (x, t) ∂RL (x, ω) pL(x, ω) = e−ωtp (x, t) dt = e−ωt dt = . R+ R+ ∂x ∂x 66 The Laplace transform of pL(x, ω) w.r.t. x is Z ∂RL (x, ω) Z pLL(u, ω) = e−ux dx = e−uxdRL (x, ω) R+ ∂x R+ Z −ux L L −ux = e R (x, ω) |R+ − R (x, ω) de R+ = uRLL(u, ω). 4.3.1 Results Based on Levy´ Subordinators ˜ For degradation with random jumps described by a Levy´ subordinator Xs(t), we n derive the explicit expressions of RX (x, t) and lifetime moments M(TX , x) in terms of Laplace transform, represented by Levy´ measures. Using the procedure similar to [112], the results are presented in Theorems 4.1 and 4.2. Theorem 4.1. For degradation with random jumps described by a Levy´ subordinator, the Laplace expression of reliability function is Z −1 LL −1 ∗ −uy RX (u, ω) = u ω + b u − (e − 1)ν (dy) , R+ where b∗ ≥ 0, ν is the Levy´ measure. ˜ Proof. Let p(x, t) be the probability density function of a Levy´ subordinator Xs(t). Based ˜ on [187], the Fokker-Planck equation for Xs(t) is ∂p(x, t) ∂p(x, t) Z ∂p(x, t) = −b + p(x − y, t) − p(x, t) + Iy∈(0,1)y ν (dy) . (4.4) ∂t ∂x R+ ∂x 67 For (4.4), we do Laplace transform of p(x, t) w.r.t. t for both sides, ωpL(x, ω) − p (x, 0) L Z L (4.5) ∂p (x, ω) L L ∂p (x, ω) = −b + p (x − y, ω) − p (x, ω) + Iy∈(0,1)y ν (dy) . ∂x R+ ∂x For (4.5), we do Laplace transform of pL(x, ω) w.r.t. x for both sides, then ωpLL(u, ω) − 1 Z LL −uy LL LL LL = −bup (u, ω) + e p (u, ω) − p (u, ω) + Iy∈(0,1)yup (u, ω) ν (dy) . R+ ∗ R Let b = b − 0