4 Prakash2020
4 Prakash2020
1 Introduction is in contrast to the other failure type, called hard, or, shock failures,
mainly caused by shocks or catastrophic events such as earthquake,
Civil infrastructure and mechanical systems such as buildings,
tsunami, hurricane, tornado, flash flood, explosion, and terrorism.
bridges, power plants, and machinery are an integral part of
While the prediction of hard failures is an equally, if not more,
modern society. Many of these systems operate under harsh operat-
important and challenging task, this paper focuses on the soft fail-
ing and loading conditions, a significant percentage nearing their
ures defined by one or multiple degradation mechanisms with pre-
end of design life. This combination of factors often leads to a
defined threshold values. To monitor the evolution of degradation
decrease in their overall reliability, imposing a substantial threat
due to various mechanisms, a system may be subjected to intermit-
to public safety and economic development. Moreover, capital
tent inservice inspections or online health monitoring, or both. The
and other resource constraints mean that replacement at the end
acquired data, collectively known as degradation data, serve as a
of their design life is either economically unjustifiable, or, socially
basis for degradation modeling and subsequent prognosis. While
unacceptable due to failure risks. To increase the reliability and
the degradation modeling focuses on the selection, calibration,
avoid catastrophic failures, proactive maintenance strategies based
and validation of competing models against the acquired degrada-
on the monitored health of system is desirable. Traditionally, relia-
tion data, prognosis is the systematic framework for assessing the
bility assessment and failure prediction are achieved using scarce
future condition or performances of an infrastructure asset, given
lifetime or failure data, which makes the reliability engineering
its present health condition. This is often done by projecting the
and system safety a “small data” problem. Due to recent advance-
degradation model to a predefined threshold. The end result of prog-
ments in wireless network, remote sensing, and edge computing
nosis is given in terms of remaining useful life (RUL), end-life,
technologies, it has become clear that the fourth industrial revolu-
health index, or the probability of failure at a given time in the
tion is taking place and the internet of things (IoT) era is fast
future.
approaching.
At the core of prognosis lies degradation modeling, which is a
way to mathematically quantify the underlying degradation mecha-
nism and subsequently predict the end-life of an unit, or the extent
1.1 General Framework of Damage Prognosis. Failures in of degradation of the unit at a specified future time [1]. The accu-
most of the engineered systems result from a gradual and irrevers- racy of future predictions depend upon several factors such as
ible accumulation of damage, which occur during their operation. choice of damage sensitive features (DSFs) also called damage indi-
This wear-and-tear type of situation, often named as soft failure, cators, the type and amount of data available, the type of degrada-
tion model used, and the measurement noise. Figure 1 illustrates
Manuscript received April 26, 2020; final manuscript received October 6, 2020; relationship between the degradation modeling and the end-life
published online November 10, 2020. Assoc. Editor: Shiro Biwa. prediction. Let Yt denote the univariate DSF observed at time t.
3 Data-Driven Models
In general, the degradation signals recorded, even from the
similar systems at same environmental and test conditions, show
significant variability in their paths. The variation from one path
to another are usually considered to be uncertainties due to measure-
ment errors, unit-to-unit differences, and internal deterioration
mechanisms [42]. Purely deterministic physics-based models
cannot always explain such uncertainties, thereby presenting diffi-
culties in estimating remaining useful life. However, data-driven
Fig. 3 Physics-based approach for degradation modeling and models do not rely on physical models and use statistical principles
prognosis (adapted from Ref. [36]) in conjunction with monitored data to construct models (such as,
Inspection number 1 2 3 4 5 6 7 8 9 10 11
Time (cycles) 0 5 10 15 20 25 30 35 40 45 50
y (C1A), Amp · h 0.99 0.97 0.95 0.95 0.92 0.90 0.87 0.86 0.85 0.83 0.82
stochastic and bivariate process models) and account for the varia- 3.1 Cumulative Damage Damage Modeling Framework.
bility in the degradation paths. The CD assumes that the infrastructure deterioration is caused by
A diagram showing the classification of data-driven degradation multiple “shocks,” and that the damage caused by each shock accu-
model is presented in Fig. 2. In the following subsections, a review mulates over time. Let Di denotes the size of the damage caused by
of cumulative damage (CD) modeling framework, directly observ- the ith shock, N(t) the number of shocks occurred till time t, and Wi
able modeling approaches, and bivariate process modeling the sequence of inter-arrival times between successive shocks.
approaches are presented. Then, the cumulative damage Y(t) can be given by
In general, the transition probability matrix is either estimated Furthermore, it can be shown that the failure time T follows a trans-
using the field data, or, based on engineering judgment. Madanat formed inverse Gaussian distribution, i.e., Λ(T) ∼ IG(ηD /ν, η2D /σ 2 ),
et al. [54] proposed a method for the estimation of infrastructure and the CDF of Λ(T ) is given by
transition probabilities from condition rating data. Estimation of
transition probabilities for bridge deterioration and subsequent fore- η2D νt
FΛ(T) (t) = Φ − 1
casting is proposed in Ref. [55] using multiple exponential hazard σ 2 t ηD
models. The main assumptions of Markov model are as follows:
2νηD η2D νt
(i) Transition probability does not change with time. + exp Φ − + 1
(ii) The sojourn time in a particular state is exponentially σ2 σ 2 t ηD
distributed. ηD
(iii) The sum of all transition probabilities out of each state into The expectation and variance of Λ(T ) are E(Λ(T)) = and
another different state must be equal to one. ν
Var(Λ(T)) = ηD σ 2 /ν3 , respectively.
The assumption that the sojourn time is exponentially distributed Wiener process and its variants such as those with time scale
is impractical for many of applications, and hence, Markov model is transformation, with measurement error, and with a drift have
not an appropriate choice in those cases. This issue can be addressed been widely applied in the literature for degradation modeling.
by using the semi-Markov model, where the time spent in a partic- Doksum and Hbyland [59] modeled the variable-stress accelerated
ular state can be given by any distribution. A hybrid degradation- life testing fatigue failure data using a WP. In their study, the accu-
based reliability model for a single-unit system whose degradation mulated fatigue decay was modeled using two separate WPs, which
is driven by a semi-Markov environment was proposed by change from one to another at a certain stress change point, t. Wang
Ref. [56]. In a recent study, semi-Markov model is used for the oxi- [60] applied a WP with random effects to analyze bridge beam
dation degradation modeling of gas turbine nozzles [57]. Despite degradation data. In their model, the unit-to-unit variability was
the use of Markov and semi-Markov model, the output of the incorporated through random effects, and the uncertainties in the
process is its state, which corresponds directly to a physical, observ- model parameters were estimated using a bootstrap method. In
able event. Nonetheless, in many practical applications, the hidden another study, Si et al. [61] used a WP-based degradation model
states of the system can only be inferred through indirect sensor with a recursive filter for RUL estimation of gyros in an inertial nav-
measurements such as vibration, strain, and displacement. igation system. They were able to update the drift coefficient of WP
∞ ∞
uv(t)
= f (y)dy = exp {−uy}dy where G(.; μΛ(t), λΛ2(t)) is the CDF of IG(μΛ(t), λΛ2(t)) and Φ(·)
ηD ηD Γ(v(t)) are standard normal CDF. Thus, for IG process there is a one-to-one
(15)
Γ(v(t), ηD u) relationship between Λ(t) and the CDF of the failure time distribu-
= tion FT(t) for fixed parameter λ and μ. Moreover, the IG process can
Γ(v(t))
easily be extended to incorporate random effects (i.e., unit-to-unit
∞
where Γ(a, y) = z=x za−1 e−z dz is the incomplete gamma function. variability) and covariates such as temperature, voltage, and pres-
The parameters of GP model can be estimated using maximum sure in the IG process model. Wang and Xu [73] applied IG
likelihood, method of moments, and Bayesian statistics [64]. A process to the laser data where both the unit-to-unit heterogeneity
Bayesian approach of parameter estimation and its update were pro- and covariate information were incorporated into the model. In
posed by Guo and Tan [65]. Ye et al. [66] obtained maximum like- their study, the EM algorithm was used to obtain the maximum like-
lihood estimates of the parameters using the EM algorithm. In their lihood estimates of the unknown parameters and bootstraping is
method, the bootstrap technique is used to construct confidence used to assess the variability in estimates. Recently, Peng et al.
intervals. [75] conducted a Bayesian analysis of inverse Gaussian process
The various extension of GP such as multivariate GP, weighted models for degradation modeling and inference. Apart from a
GP, and non-stationary GP are also employed for degradation mod- simple IG process model, they also investigated three IG process
eling. van Noortwijk [64] presents an exhaustive survey of models with random effects using the Bayesian method.
GP-based degradation modeling for civil engineering infrastructure.
A stochastic GP model was used for building deterioration by Edir- 3.2.5 Time Series Models. Time series models such as auto-
isinghe et al. [67], for creep of concrete by Cinlar et al. [68], for regressive (AR), auto-regressive moving average (ARMA), and
fatigue crack growth by Lawless and Crowder [69], and for thinning auto-regressive integrated moving average (ARIMA) models have
caused by corrosion by van Noortwijk et al. [70]. Grall et al. [71] been widely used for short-term forecast [76]. The fundamental
used a GP to analytically model the deterioration of a continuously idea of these models is to represent the future DSF values as a
deteriorating single-unit system and proposed an optimum preven- linear function of past DSF observations and random errors. In
tive replacement policy structure. Liao et al. [72] presented a general, an ARMA(p,q) model consists of p autoregressive (AR)
condition-based maintenance model for continuously degrading and q moving average (MA) parameters and can be written as
systems under continuous monitoring. follows [76]:
In past, GP has been widely used for degradation modeling due to
its physical interpretation and rich mathematical structure. ϕ(B)(Yt − μ) = θ(B)at (19)
Fig. 8 Architecture of a Hidden Markov model Fig. 9 Schematic of a two-phase model with a change-point
Then, a general multi-phase degradation model can be written as A two-phase degradation modeling approach was used for rolling
follows: elements bearings by Gebraeel [117]. In the aforementioned work,
the authors developed a two-phase degradation model using the
Yt = [Y0 + Y1 (t)]I(0,λ1 ) (t) + [Yλ1 + Y2 (t − λ1 )]I(λ1 ,λ2 ) (t) root-mean-square value of acceleration, and they integrated real-time
(20)
+ [Yλ(n−1) + Yn (t − λ(n−1) )]Iλ(n−1,∞) ) (t) condition data in a Bayesian framework. In their model, the first
phase consists of a relatively constant signal representing the
where I(a,b)(t) is the indicator function, which is unity in interval [a, period before damage initiation, followed by a second phase—
b] and zero elsewhere. The simplest case of a two-phase degrada- monotonically increasing—representing damage progression.
tion model can be written as follows: Using this approach, they arrived at a closed-form solution to
predict and update the RUL. Change point detection along with
Yt = [Y0 + Y1 (t)]I(0,λ) (t) + [Yλ + Y2 (t − λ)]I(λ,∞) (t) (21) RUL estimation was studied by Chen and Tsui [118] and Prakash
For example, if the functional form is linear in both phases, the et al. [119]. The authors in these studies proposed a Bayesian
model can be written as follows: change point detection approach for two-phase degradation model-
ing and RUL predictions.
θ1 + θ2 t + ϵ1 if t ≤ λ Two-phase degradation modeling with two different functional
Yt = (22)
β1 + β2 t + ϵ2 if t > λ forms in the two phases has been undertaken by Wang et al. [120]
for real-time reliability evaluation of LCDs, in which the initial
where λ is the change point location, ϵ1, ϵ2 are the additive errors, sharply increasing phase was modeled using a Gamma process,
which are assumed to be i.i.d and normally distributed with mean while the latter using a Weiner process. Feng et al. [121] used a multi-
0 and variances σ 2 and τ 2, respectively. Let θ1 = {θ1, θ2, σ 2} and phase Weiner degradation model for storage life prediction for a
θ2 = {β1, β2, τ 2} then the model parameters Θ = {θ1, θ2, λ} can high-performance capacitor. It was found in this study that ignoring
be estimated either using a maximum likelihood or Bayesian change points that exist in the degradation path could lead to signifi-
approach. According to the Bayes theorem, the posterior distribu- cant prediction uncertainty in the storage life of a capacitor.
tion of the parameters, p(Θ|Y) can be given as follows:
π(Θ)L(Y|Θ) 3.4 Example: Data-Driven Models. The problem of battery
p(Θ|Y) = ∝ π(Θ)L(Y|Θ) (23)
p(Θ)L(Y|Θ)dΘ degradation as discussed in Sec. 2.1 is further undertaken using a
data-driven model. A regression-based model, namely, Gaussian
where π(Θ) is prior distribution obtained from the past degradation process model has arbitrarily been chosen for this purpose. The
and L(Y|Θ) is the likelihood of observed data Y = {Y1, Y2, …, Yt}. Gaussian process regression (GPR) has widely been used in data-
For the degradation model in Eq. (22), the likelihood function driven prognostics. In contrast to linear regression, which
L(Y|Θ) is written as follows: assumes that the error terms are independent and identically distrib-
uted (i.i.d.), the GPR assumes correlated errors and results in rela-
L(Y|Θ) = ϕ(Yt ; θ1 + θ2 t, σ 2 ) ϕ(Yt ; β1 + β2 t, τ2 ) (24)
t≤λ t>λ
tively accurate predictions.
The results of GPR simulations for battery degradation are shown
where ϕ(Yt; θ1 + θ2 t, σ ) denotes a normal probability density func-
2
in Figs. 10 and 11, respectively.
tion with mean θ1 + θ2 t, variance σ 2 and evaluated at Yt. The main Figure 10 shows the predicted median and 90% prediction inter-
assumptions of change point model is as follows: (i) the condition val similar to physics-based model. The threshold that is to 70% of
of the infrastructure deteriorates with operating time and the level initial C1A value and true degradation path are also shown in the
of deterioration can be observed at periodic intervals; (ii) the infra- figure. It can be seen from the figure that the predicted degradation
structure being monitored comes from a population of similar units, path is well in agreement with the actual degradation path. Finally, a
and each of which exhibits the same degradation form; (iii) the distri- histogram of predicted RUL at 50 cycles is plotted in Fig. 11.
bution of the random term across the population of devices is known; Note that, the actual RUL for the battery under consideration is
and (iv) a change point occurs only due to the damage. 68 cycles, while the 95% prediction is 36.2 cycles. The accuracy
of prediction will further increase as the more data will be included (e.g., failure time is less <2 year) when such rules are known to
in this model. the human experts. In formulation of expert systems, a statement
can be either true or false; data can thus be included or excluded
from a set. However, it is not always feasible to define sets and asso-
4 Knowledge-Based Models ciated membership so precisely. Fuzzy logic is one of the first exten-
sions of expert system in which a probability is associated with each
Knowledge-based approaches are generally employed when
rule in place of simply using true or false rules to represent the
accurate mathematical models are difficult to build due to real-
knowledge.
world complexity or due to various limitations of using model-
based approaches. A knowledge-based system is essentially
composed of two sub-systems: the knowledge base and the infer- 4.2 Fuzzy Logic. Fuzzy logic is a problem-solving methodol-
ence engine. Two typical examples of knowledge-based systems ogy that provides a simple way to arrive at a definite conclusion
are expert system and fuzzy logic. based upon vague, ambiguous, imprecise, noisy, or missing input
information. Fuzzy logic is widely used in various systems ranging
from simple, small, embedded micro-controllers to large, networked
4.1 Expert System. An expert system is a computer that emu-
PC or workstation-based data acquisition and control systems. In
lates the decision-making ability of a human expert [122]. It gener-
other words, fuzzy set theory allows a partial set membership
ally consists of a knowledge base containing accumulated
based on a variable’s “degree of truth.” As in expert systems, fuzzy
experience from the experts in the field and a rule base for applying
logic systems use simple, empirically derived, IF-THEN rules to
that knowledge to particular problems known to the software
solve problems. These rules are descriptive but unlike expert
system. In other words, expert systems are designed to solve
systems, they are intentionally imprecise. A typical fuzzy process
complex problems by reasoning through bodies of knowledge,
logic statement may look like “IF (process is too hot) AND
represented mainly as IF-THEN statements rather than through con-
(process is heating rapidly) THEN (cool the process quickly).”
ventional procedural code. The process of building an expert system
The fuzzy logic-based method has been successfully imple-
involves knowledge acquisition, knowledge representation, and the
mented for several applications such as dam health monitoring
verification and validation of models [122]. Once an expert system
[124], prediction of end-life of boiler tubes [125], active control
is built, it can be used for solving problems which are normally
of bridge structures [126], and active control of high rise buildings
solved by human specialist. The output of an expert system is gen-
[127]. Interested readers for the several engineering applications of
erally understandable by non-specialist users, and logic for a partic-
fuzzy logic are directed to the book by Ross [128]. In the following
ular result can be established. Expert system has been widely
section, a numerical example is presented to illustrate the
applied for various purposes, such as for diagnosis, prognosis, inter-
knowledge-based model.
preting, monitoring industrial systems, and predictive maintenance
activities [31]. Liao [123] presented an exhaustive review of appli-
cations of expert systems between 1995 and 2004. 4.3 Example: Knowledge-Based Model. A numerical
Despite of its wide applications, the expert system approach has example using Fuzzy logic to estimate the residual life is presented
some drawbacks. First, the problem of “combinatorial explosion,” here. The details of the algorithm can be found in Ref. [123]. Let us
which often occurs when the number of rules (i.e., if-then state- consider the data given in Table 3 for two sample units of an infra-
ment) increases dramatically and causes computational issues. structure, which includes observed degradation index (x1), thresh-
Second, the performance of an expert system depends upon the old (xs), observed degradation rate per year (x2), and elapsed time
experts that develops the laws of reasoning. Finally, it is not possi- (y). Using this dataset, it is desired to obtain the residual life of
ble to make continuous variable predictions using expert system infrastructure facility.
because the output is determined by a discrete set of rules. This In Table 3, each sample is observed data from different elements
limits their ability to estimate remaining useful life. However, an of the facility. In fuzzy logic, the membership function is defined to
expert system can provide an approximate output in general terms represent practitioner’s knowledge, and here, it is assumed
Fig. 12 Result of numerical example (a) values of the membership function and (b) estimated remaining useful life
Fig. 13 Installed Instruments in TK dam (a) strain Gauge and (b) seismic tri-axial MEMS accelerometer and recorder
dynamic loads such as earthquakes and hurricanes cannot be HM4: data-driven model + physics-based model; and (v) HM5:
completely described using physics-based models or purely stochas- experience-based model + data-driven model + physics-based model.
tic processes. Similarly, remaining useful life of a mechanical com- In HM1 approaches, the domain knowledge is integrated with
ponent such as a cutting tool or rolling element bearing cannot be continuous SHM data to build a model. The domain knowledge pro-
accurately predicted only by using a data-driven techniques [129]. vides an insight into the system’s fault states, while data-driven
In such cases, a hybrid degradation model is particularly useful, models are used to refine the rules generated from expert knowledge
where the partially understood degradation mechanism, and SHM and to perform the actual RUL prediction. Using this approach,
data acquired from the system can be utilized to build a model. Garga et al. [130] presented an automated reasoning method that
The fundamental assumption of hybrid model is that the use of uses rule-based expert system, and neural network for the remaining
various models helps to model the degradation process efficiently. useful life prediction of an industrial gear-box. The HM2 approaches
There are several ways by which these models can be combined to integrate experience-based models and physics-based models, in
build a more efficient degradation model. In general, Liao and which the output of the experience-based model is often used to
Köttig [35] categorized the various hybrid approaches into following enhance the physics-based model. Byington et al. [131] presented
five types: (i) HM1: experience-based model + data-driven model; (ii) a methodology for the RUL prediction of aircraft actuator compo-
HM2: experience-based model + physics-based model; (iii) HM3: nents, where the fuzzy logic was used to quantify the level of
data-driven model of type + data-driven model of another type; (iv) damage (damage index), and Kalman filtering was utilized to
(a) (b)
Fig. 15 Dam performance model using hybrid degradation model (a) X bar and (b) range chart for principal strains
Physics-based
[36] A general framework Remaining useful life
[30] A review Rolling element bearing
[8] Deterioration modeling of structural systems Corrosion and fatigue
[40] Fick’s diffusion law Concrete structures
Data-driven
Bivariate process [91] Adaptive Bayesian decision Replacement policy
[87] Objected-oriented Bayesian networks Time series
[92] Bayesian Network concrete bridges
[93] Kalman filter Conditional residual life
[94] Kalman filter Aircraft power generators
[95,96] Particle filter Fault indicator, fatigue degradation
[97] Stochastic filter Remaining useful life
Cumulative [44–48] Cumulative degradation Ball bearings, composite laminates, fatigue life, fiber-reinforced
damage plastics, and railway wheel damage.
[49] Dependent competing failure Reliability model
[50] Degradation and shock loads Multi-component system reliability model
[51] Dependent competing risk model System subject to shocks
[52] Gradual and shock deterioration processes Life-cycle analysis
Markov [100] Hidden Markov model Diagnosis and prognosis
processes
[138,139] Mixture of Gaussian HMM RUL estimation
[105,140,141] Hidden semi-Markov model Diagnosis & prognosis of pumps
[101,142,143] Hidden Markov model Rolling element bearing
[144–146] Markov model Bridge deck systems
[147] Semi-Markov model Bridge deterioration modeling
Gamma process [64] Theoretical aspects Inspection and maintenance
[67,148,149] Time-dependent reliability Deteriorating structures
[69] Covariates and random effects Fatigue-crack-growth
[20,150,151] Parameters and RUL estimation Corrosion of nuclear pipes
[68] Calculation of confidence limits Creep of concrete
[152] Bivariate degradation Fatigue crack
[153] With random effects Laser degradation
[71,72,154] Condition based maintenance Numerical example
[155] Integration of SHM data Creep deformations of bridges
[148] Temporal variability and uncertainties Concrete sewer pipes
[156] Aging and shock events Bilinear structural systems
Weiner process [62] Time-scale transformation Heating cable
[63] With measurement error Wear of magnetic heads
[60] With random effects bridge beam degradation
[61] RUL using recursive filter Gyros used for navigation
[157] Bivariate Weiner process Aluminum reduction cells
[158] Time correlated structure Resistors and sliding metal
[159] Real-time reliability evaluation Capacitance loss over time
Inverse Gaussian [160] With random-drift Optimal CBM policy
[60] Maximum likelihood estimation Laser devices
[74] Random effects, explanatory variables Laser devices
[161] Bayesian method using inspection data Corrosion growth modeling
[75] A Bayesian perspective GaAs laser degradation data
Other models [117–119] Two-phase degradation Rolling element bearing
[120] Two-phase Wiener and gamma Liquid coupling diode
[121] Multiphase Wiener High storage capacitor
[77–79] Time series model Rotating machinery
[76,80] Long memory models Long-term time series data
[106] Neural network model Rolling contact thrust bearings
[108] Neural network model Highway culverts
[109] Recurrent neural network model and support Pavement performance
vector classifier
[110] Elman neural network model Bridge elements
[107,111– Various machine learning techniques Bridges, water and sewer pipes
116]
Knowledge-based
Expert system [31,122,123] Diagnosis, prognosis, and monitoring Industrial systems
Fuzzy logic [124– Monitoring, diagnosis, and prediction Dam, boiler, bridge high-rise building
127,162]
Hybrid-models
[129] Hybrid approach using Wiener process Cutting tool
[35] Categorization of hybrid approaches Five types
[130] Expert system and neural network Gear box
[132] Time series and statistical process control Multi-storey buildings
Advantages Drawbacks
Physics-based approaches
• Among all the approaches, the physics-based approach provides most • An detailed knowledge of system and underlying degradation process is
accurate and precise RUL estimate required
• For physics-based models, an analytical expression for degradation • These approaches make several assumptions about system to formulated
phenomenon, and a closed form expression for end-life can be derived, the problem, which may not be the case for a unit operating in the field
which can be used for effective maintenance planning
• In this approach, the effect of different variables on degradation process can • These approaches consider degradation as a deterministic process, and
be investigated. Hence, by controlling the some of the variables system’s doesn’t account for uncertainties present due to measurement errors,
life can be increased unit-to-unit variability, and material heterogeneity
• These models are formulated using the fundamental engineering concepts • It is not always possible to formulate the degradation process in terms of
and help to physically understand the underlying degradation mechanism mathematical equations, since the underlying process is complex and
and different modes of system failure unknown
• The physics-based model require relatively less data for parameters • In some cases, it requires finite element modeling, which is
estimation compared to other models such as data-driven model or computationally expensive, if the system is complex
experience- based model
Knowledge-based approaches
Expert system
• Expert system is simple to develop and easy to understand • It depends upon the quality of input knowledge from the domain experts
• Expert system rules can be easily reviewed, and even edited by domain • The model requires a significant number of rules that has to be updated
experts rather than IT experts with time for accurate results
Fuzzy logic
• Fuzzy logic requires relatively fewer rules compared to Expert system • It requires domain knowledge to develop the fuzzy rules
• Input can be imprecise since, the rules are given in term of probability • Assignment of weight to the input variables is difficult decision
• Both ES and FL are easy to maintain. Also, a rapid prototype can be created • It hard to obtain the domain knowledge and subsequently convert them
with a few rules in days rather than the months or years into rules
Data-driven approaches
General comment
• This approach is general and can be used to model any kind of degradation • Many researchers considers this approach as a black-box, which totally
process, once the degradation data are available relies on the data without much knowing about the underlying physics
• With the choice of different functional form, it is possible to generate a • Domain knowledge can be integrated only if the parameters are estimated
wide verity of data-driven degradation models, e.g., random variable, in Bayesian manner
stochastic, Markov and time series
• The various machine learning algorithms such as SVM, NN, clustering, • Most of the data-driven methods require training data collected for all
GMM, and SPC can directly be applied to the data possible operating and loading conditions for an accurate prognosis
Random variable model
• A close form expression for the CDF of failure time, T can be derived • A fixed functional form for a particular degradation process does account
for the variations in environmental and loading conditions
• With the change of functional form a large number of degradation • This model does not account for temporal uncertainty, i.e., uncertainty
processes can modeled associated with the evolution or progression of deterioration over time
Stochastic models:
Markov model
• The mathematics of Markov model is well established and can be used to • It assumes a single monotonic and non-temporal failure degradation
model various the failure scenarios pattern
• Markov model can be applied with the incomplete data sets. Also, the
knowledge of underlying degradation process is not required
• Markov model is computationally efficient, once developed. The calibrated • If time spent in a given state is not exponentially distributed then more
model can be used to predict the RUL along with its confidence interval complex semi-Markov model is required
Hidden Markov model and hidden semi-Markov model
• HMM models the degradation of a system in phases (i.e., hidden states), so • Data corresponding to various states of system are required to train the
the knowledge of continuous degradation path is not necessary HMM model, which increases rapidly with increasing number of hidden
states
• Both HMM and HSMM uses indirect measurements (e.g., vibration, • The suitable computational tools (or software) for HMM and HSMM are
acoustic etc.) obtained from the system not available
• The details of failure progression mechanism is not required • Both are computationally intensive particularly in case if modeling
requires a large number of hidden states
Lifetime data-based approaches
• Lifetime data based approach is most popular among the reliability • This approach does not account for the variation in operating and loading
engineer due to its simplicity compared to the other approaches condition of an infrastructure
Advantages Drawbacks
• This approach is computationally in-expensive. The calculations can be • Very often, for long-life and high reliable infrastructure lifetime data
performed even on a hand-held calculator remain unavailable
• This method particularly useful in cases where the condition monitoring • In sample where the lifetime data are relatively sparse, reliability estimates
data is difficult to obtain can be overly pessimistic
• This is a population-based approach particularly useful for maintenance • This approach does not focus on the degradation of a particular unit
planning of low cost units
Hazard model
• It integrates the lifetime data and condition monitoring data for reliability • All relevant covariates must be included in the model
assessment
• Softwares are available for computations • Historical lifetime failure data are required
Time series model
• Advanced ARMA model can be applied for non-stationary data • Basic ARMA models assumes that the data are stationary, which is often
not the case
• Lifetime failure data is not required for time series model • Measurement noise present in the signal can result in an inappropriate
model-order selection
• Like other data-driven models time-series approach also does require full • A significant amount of data is required for model parameters estimation
understanding of degradation mechanism
• The various software and computational tools are freely available for fitting • Unlike proportional hazard model, TS model doesnot include the lifetime
a time-series model data and covariate information in analysis
• Time series model can provide accurate short-time forecast • Long-time forecasts using TS model is less reliable
Change point-based model
• Model predicts more accurate RUL when degradation is in the second • Detection of a change point in real-time is a challenging task and can often
phase result in a false alarm
• In this model, prior knowledge can be integrated for improved maintenance • For many long life infrastructure, the prior knowledge about the change
decision of a monitored unit point location is not unavailable
Hybrid approaches
• This approach is useful where limited data is available and some domain • Often two different model one for estimation, and other for prediction is
knowledge partial degradation mechanism is understood required
• A large class of models are available for this approach • In some cases, it may be challenging to integrate the two different types of
model
predict the progression of the damage. In HM3 categories, two differ- to detect any abnormal behavior in the dam responses in real-time. The
ent data-driven model can be fused to come up with a more efficient proposed method is applied to the strain and displacement data of
hybrid model. Here, the first model can be used to estimate while the Tsankov Kamak dam located in Bulgaria. The readers can refer the pre-
second can be used to predict the system state. Recently, Prakash and vious study done by the authors [137] for the details of dam, sensor
Narasimhan [132] used combined time series model in conjunction placement, and data acquisition.
with statistical process control chart for remaining useful life predic- To monitor the health of this dam, several types of instruments
tion of multi-storey buildings. Among all hybrid models, HM4-type were installed across the dam, and data were collected using an
hybrid degradation model is widely applied in the literature. For extensive data collection system. Two of such sensors strain
example, Xu et al. [133] developed a physics–statistics-based degra- gauge and seismic tri-axial accelerometer are shown in Fig. 13.
dation model for reliability predictions and illustrated for data The six strain components (after subtracting the zero-stress strain
obtained from accelerometers for accelerated degradation test. Liu measurements) corresponding to left and right block of crest (top
et al. [134] integrated data-driven prognostics together with a middle block of dam) are shown in Fig. 14, which are used to cal-
physics-based particle filter for system state prediction; Mohanty culate the principal strains and their orientations. In Fig. 14, the sea-
et al. [135] combined a physics-based state space model and a data- sonal trends in strain measurements for both the blocks can clearly
driven kernel-based Gaussian process regression model for fatigue be seen. Moreover, a decreasing trend with time which can be con-
crack growth modeling in metallic alloys. Recently, Paris crack sidered as irreversible is also present for XZ, YZ, and ZZ directions.
growth law in conjunction with the SHM data is utilized to model To model both the seasonality and trend in the data,a hybrid
the fatigue-induced damage in a recursive Bayesian manner by HSTT model is considered as follows:
Gobbato et al. [136]. Finally, all the three approaches, i.e.,
experience-based, data-driven, and physics-based, can be utilized y = yH + ys + yT + yt + ϵ (26)
to built a HM5-type hybrid model. Hybrid approaches in the HM5
type is impractical to implement due to the difficulty that might be where y is the dam response; yH, ys, yT, and yt are the contributions
encountered by each type of model. due to hydrostatic load, seasonal effect, variation in temperature,
and age-related deterioration were applied. The error term ϵ is
assumed to be independent and identically distributed. The follow-
ing functional forms for yH, ys, yt are considered:
5.1 An Application: Hybrid Model. The first author of this
paper has successfully implemented a hybrid model (HM3 type) for
N
M
dam health monitoring. This hybrid prediction model utilizes hydro- yH (H) = ai H i ; yT (T) = bi Tii ;
static, seasonal, temperature and time response data collected from i=1 i=1
the dam and predicts the dam performance during initial service life.
The dam responses such as strain and crest displacement are modeled yt (t) = c1 t + c2 exp(−t) (27)
using recursive least squares while the thresholds are set using univari- ys (θ) = d1 sin(θ) + d2 cos(θ) + d3 sin(θ) cos(θ);
ate and multivariate control charts, based on residuals between the mea-
sured and predicted responses. Furthermore, both the model parameters 2πt
with θ =
and thresholds are updated as more data becomes available, which helps 24 × 365