IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 48, NO.

9, SEPTEMBER 2018 1637

A Model-Based Hybrid Approach for Circuit

Breaker Prognostics Encompassing Dynamic
Reliability and Uncertainty
Jose Ignacio Aizpurua, Victoria M. Catterson, Senior Member, IEEE,
Ibrahim F. Abdulhadi, and Maria Segovia Garcia

Abstract—Prognostics predictions estimate the remaining Q(z, ·) Probability of transition from z in E × E.

useful life (RUL) of assets. This information enables the imple- tp Prognostics prediction time instant.
mentation of condition-based maintenance strategies by schedul- RU Ltp Remaining useful life at tp .
ing intervention when failure is imminent. Circuit breakers (CBs)
are key assets for the correct operation of the power network,
U Nonfailed state of the system.
fulfilling both a protection and a network reconfiguration role. U Failed state of the system.
Certain breakers will perform switching on a deterministic sched- L(θ |x) Likelihood of a set of parameter values, θ ,
ule, while operating stochastically in response to network faults. given outcomes x.
Both types of operation increase wear on the main contact, with μy1 ,...,yn Conditional distribution of Z(t) given available
high fault currents leading to more rapid aging. This paper observations {y1 , . . . , yn }.
presents a hybrid approach for prognostics of CBs, which inte-
grates deterministic and stochastic operation through piecewise Rz (t) Reliability of the system at time t when the
deterministic Markov processes. The main contributions of this initial state is z.
paper are: 1) the integration of hybrid prognostics models with f (t) Probability density function (PDF).
dynamic reliability concepts for a more accurate RUL forecasting (t) Stress signal.
and 2) the uncertain failure threshold modeling to integrate and L(t) Aging variable.
propagate uncertain failure evaluation levels in the prognostics
estimation process. Results show the effect of dynamic operation
β Shape parameter of the Weibull distribution.
conditions on prognostics predictions and confirm the potential η Scale parameter of the Weibull distribution.
for its use within a condition-based maintenance strategy. fU PDF of the failure threshold.
Index Terms—Circuit breaker (CB), dynamic reliability, hybrid
rU Randomly sampled failure threshold.
model, physics of failure, prognostics, uncertainty. tjump Jumping time or failure occurrence time.
ttrip Instant in which the circuit breaker (CB)
receives open signal.
topen Instant in which the CB reports being open.
N OTATION tswitch CB switching operation start time.
SF6 Sulphur hexafluoride. tend_switch CB switching operation finish time.
I2t Cumulative degradation (CD). tclear CB fault clearing operation open time.
Z(t) Piecewise deterministic Markov iRMS Current at the time of operation.
process (PDMP). ifault Fault current.
(x, t) Deterministic motion at position x at time t. μrep Constant repair rate.
Xn State of the system after the jump n. D Total number of network fault shocks.
Tn Time instant n. k1 Normalizing constant of the cumulative
N(Xn , ·, ·) Renewal Markov kernel. shocks.
s Time instant in the process state space. k2 Normalizing constant of the delay.
E Probability space. sshock Shock occurrence indicator variable.

Manuscript received January 23, 2017; accepted March 16, 2017. Date of I. I NTRODUCTION
publication April 12, 2017; date of current version August 16, 2018. This work
HE increasing complexity of power networks includ-
was supported by the Engineering and Physical Sciences Research Council
under Grant EP/M008320/1.
J. I. Aizpurua and V. M. Catterson are with the Institute of Energy
T ing smart grid technologies requires efficient mainte-
nance solutions to ensure the safe and cost-effective opera-
and Environment, University of Strathclyde, Glasgow, U.K. (e-mail:
[email protected]; [email protected]).
tion of its constituent assets [1]. Many of these assets are
I. F. Abdulhadi and M. S. Garcia are with the Power Networks reaching the end of their life and condition-monitoring and
Demonstration Centre, University of Strathclyde, Glasgow, U.K. (e-mail: life extension techniques are emerging as feasible technolo-
[email protected]; [email protected]). gies for enhanced maintenance planning of these systems
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. (e.g., [2], [3]). The implementation of these approaches
Digital Object Identifier 10.1109/TSMC.2017.2685346 presents significant challenges for each specific asset.
2168-2216 c 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
1638 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 48, NO. 9, SEPTEMBER 2018

Fig. 1. Network requiring the operation of CBs.

Prognostics is the ability to acquire knowledge about events

before they actually occur [4]. In engineering, failure prog-
nostics aim to predict the RUL of a component after a fault
is diagnosed, i.e., the fault-to-failure progression of an asset
(e.g., [2], [4], [5]).
Successful implementations of prognostic applications pro-
vide benefits for asset maintenance planning through the
extension of the useful life and reduction of maintenance
actions. Consequently, condition-based asset maintenance poli-
cies which integrate prognostics predictions when planning a
maintenance schedule, can result in cost-effective operation of
assets (e.g., [6]–[8]).
However, some assets prove particularly challenging when
predicting failure, due to their operating regime incorporating
Fig. 2. Cross section of a CB derived from [9].
both deterministic and stochastic behavior. This paper intro-
duces a novel model-based hybrid approach to prognostics for
such assets, with a particular focus on power network CBs.
2) When the CB begins to open as shown Fig. 2, there is no
arc as the current keeps flowing from the upper terminal
to the moving parts through the fixed and moving arc-
A. Circuit Breaker Operation ing contacts (see current in Fig. 2, dashed-line filled
CBs are critical elements for the correct operation of the arrowheads).
power network. Their objective is to interrupt current flow, 3) During the opening of the moving arcing contacts
and they operate either: (Fig. 2, solid-line empty arrowheads) the current keeps
1) in response to a network fault—fault operation; flowing due to the arc generated between the fixed and
2) or to reconfigure the network (e.g., for switching in moving contacts.
response to daily load changes)—switching operation. 4) The arc is quenched when the CB is in the fully open
Consider Fig. 1, where there are two paths for the generator position. In a high current situation (e.g., network fault)
to supply the load (along line 1, and along lines 2 and 3). At the puff of SF6 may be needed to quench the arc.
times of peak load, the current will flow along both paths. As Each operation places wear on the main contact of the CB,
the loading drops to a level where a single path has sufficient which is proportional to the heating in the contact caused by
capacity for the current, CBs A and B will operate to switch the arc. This wear is referred to as I 2 t, meaning the square
out one path. Traditionally, there is one peak in the load per of the current, I, multiplied by the time the arc is present,
day, so switching operations occur once per day. t [11]. Every time the CB operates, more wear accumulates
However, if a fault occurs on line 1, CBs A and B should on the main contact. The total cumulative wear increases
operate to clear the fault and protect the rest of the network. monotonically by the I 2 t of each operation.
A failure to operate will have significant safety and financial Switching operations will tend to occur every day when the
consequences, since the network fault will need to be cleared current is at the same relatively low level. Since a lower current
by a more distant CB, which takes longer and removes more can be broken more easily, the arc on the main contact will
customers from service. Accordingly, CBs are required to have not exist for long, and I, t, and correspondingly I 2 t will be
a very high availability. relatively low. Switching behavior is relatively deterministic,
Fig. 2 shows the cross section of a SF6 CB comprised of as it occurs under the same conditions each day.
main and parallel arcing contact systems. The main contact However, whenever a network fault occurs, the current can
conducts the current between the fixed and moving parts of increase by one or two orders of magnitude, and it takes longer
the breaker and the arcing contact takes care of the arcing to quench the subsequent arc. This situation places far higher
during separation. The nozzle guides SF6 gas to efficiently I 2 t wear on the breaker. Different types of network fault,
quench the arc. such as phase-to-phase faults, single phase-to-ground faults,
The CB operates as follows [9], [10]. and three phase-to-ground faults, will cause different levels of
1) When the CB is in the closed position, the current flows fault current to flow. Since the type and severity of a fault
from the fixed main contact to the moving main contact is stochastic, the value of I 2 t is also stochastic under fault
(Fig. 2, dashed-line empty arrowheads). conditions.

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
AIZPURUA et al.: MODEL-BASED HYBRID APPROACH 1639

B. Circuit Breaker Failure Modes and Maintenance

As described above, operation of a CB places cumulative
wear on the main contact due to the presence of the arc. If not
maintained, the CB would eventually reach a point of failure.
Fail-to-open and fail-to-close are the main failure modes of
CBs and excessive contact wear is a critical event that directly
causes the occurrence of these failure modes [12], [13]. For
more detailed failure models of CBs refer to [13]–[15].
In order to ensure current can be broken in a timely manner,
breaker maintenance tends to be scheduled periodically based
on the number of operations (see Section II-A). However, this
Fig. 3. CB maintenance versus current interrupted [18].
preventive strategy does not take into account the dynamics of
the system such as the different aging effects of fault clear-
ing operations versus switching operations. Therefore, it may analysis [17]. This way, it is possible to update the static
result in a conservative (hence expensive) maintenance strategy failure model according to changes in the system operation,
which replaces assets even with a significant RUL. and accordingly replicate and predict the future behavior of
So as to implement intelligent maintenance policies, diag- the asset more accurately.
nostics and prediction steps are needed to reveal the actual On the other hand, predicting the point in time at which
state of the CB and predict its RUL, respectively [5]. Although the failure of an asset will occur requires the specification
the implementation of diagnostics techniques has been well of a specific failure threshold. However, the definition of a
studied for CBs (see Section II-A for CB diagnostics exam- deterministic failure threshold can be a difficult task for real
ples), prognostics models for CBs are scarce. Accordingly, so applications. For a population of identical assets exposed to the
as to advance in the RUL prediction of CBs, the main focus same loading, some will fail sooner than others, and population
of this paper is on the prediction step. failure time can be expected to conform to some distribu-
In order to make prognostic predictions, it is necessary to tion. In this paper, the specification of the failure threshold
create a model which replicates the future behavior of the CB is included in the prognostics model as a PDF, and we propa-
and estimates the RUL from a specific time instant onward. gate the effect of the uncertainty level to the final prognostic
To this end, it is crucial to select a problem-specific prognos- estimation.
tic prediction model depending on the system requirements, Accordingly, the main contribution of this paper is the
available engineering resources (run-to-failure data or physics- integration of a model-based hybrid prognostics model with
of-failure engineering knowledge), and analyzed failure mode. dynamic reliability concepts for more accurate prognostics
See [16] for a prognostics model-selection process according predictions of CBs. The second contribution is the integra-
to design requirements. tion of an uncertain failure threshold level in the prognostics
The main contact wear occurs due to both deterministic estimation.
switching operation and stochastic network faults. Traditional
mathematical processes such as Gamma or Wiener processes
are not well suited for this problem because they model non- D. Organization
monotonic and continuous monotonic degradation patterns, Section II discusses this paper in the context of other rele-
respectively [16]. vant work. Section III presents the generic proposed approach,
Model-based hybrid prognostic approaches have the capa- while Section IV outlines its application to CBs. Section V
bility to combine continuous and discrete time behavior of the applies the proposed approach to a case study CB. Finally,
system. This combination is suited to applications which need Section VI draws the conclusions and identifies future research
to consider a deterministic system behavior (e.g., switching- challenges.
related degradation) where the deterioration is increased at
discrete time instants (e.g., network faults). II. R ELEVANT W ORK
A. Circuit Breaker Condition Monitoring
C. Contributions An average size utility can have thousands of CBs compli-
It is possible to modify existing hybrid prognostics cating their inspection and maintenance. Accordingly, electric
prediction approaches (see Section II-B) in order to improve utilities have traditionally performed preventive time-periodic
the RUL prediction accuracy. maintenance programs to assess the condition of CBs. The
On the one hand, in order to model the future behavior of CB maintenance tends to be scheduled periodically based on
the asset, the discrete behavior has been traditionally specified number of operations [9], [10], [18]. Fig. 3 shows an example
with an a priori determined stochastic model. It is possible to of operation-based maintenance program, where the remaining
update this static model with deterministic operational signals number of operations can be inferred from the current inter-
(e.g., network loading) that modify the stochastic behavior rupted. For instance, if the current interrupted remains lower
according to the usage conditions. Dynamic reliability tech- than 2500 A the estimated number of operations are 10 000,
niques are capable of handling component interactions and however, it can decrease down to 100 operations when the
they constitute a more realistic system modeling for reliability current interrupted is 25 kA.

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
1640 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 48, NO. 9, SEPTEMBER 2018

The CB operation can be tracked through failure precur- the random evolution of the system to the stochastic shocks
sor variables indicating degradation, such as SF6 density, trip using continuous-time Markov chains.
coil current profile or I 2 t CD. These variables have been PDMPs have also been used to solve dynamic relia-
used to propose a number of CB data-mining and diag- bility problems (e.g., [30], [31]). Instead of considering
nostics approaches. The seminal work in [19] presented an a priori established failure specifications of components, these
automated CB diagnostics system based on control signals approaches add the possibility to specify components with a
implemented through signal processing and expert system variable failure rate which depends on the operating conditions
techniques. Similarly, a data-mining process was proposed of the system. This community has also started to create design
in [20] to derive data-driven diagnostic indicators from fea- tools to create hybrid models from user-friendly specifications,
tures of the trip coil current profile. This approach provides e.g., implementing PDMPs in Python [32] or linking reliabil-
CB maintenance decision support combining classification ity analysis and multiphysics specification tools [33]. There
techniques and expert interpretation. are other techniques which can also be used to solve dynamic
Data-mining techniques for CB’s condition analysis has also reliability problems such as dynamic Bayesian networks [34],
been applied more recently. Lin et al. [21] applied fuzzy set stochastic activity networks [35], or fluid stochastic Petri
theory to analyze and integrate different CB operation and nets [36].
lifetime indices. Moreover, evidential reasoning generates the Apart from PDMP models, there are other hybrid prog-
overall assessment of the CB condition. A framework to assess nostics approaches. A new formalism called hybrid particle
the CB condition using the control signal waveforms was Petri nets combines Petri nets with particle filtering for health
presented in [22]. The authors used classification techniques monitoring studies [37]. The model includes deterministic and
to categorize the health of the CBs with qualitative (normal, stochastic properties and it is mainly focused on diagnostics.
alarm and emergency) and probabilistic quantitative indices. Recently, Daigle et al. [27] proposed a generic model-
The implementation of prognostics models is not as well based hybrid prognostics architecture inspired by hybrid bond
studied as diagnostics techniques. As pointed out recently graphs. The continuous dynamics are defined with a model
in [23], the ability to predict the aging of CBs is not fully comprised of components, which in turn have variables with
developed, as they do not have a clearly defined physics-of- constraints. The discrete dynamics are defined with finite state
failure equation model. To the best of authors’ knowledge, machines that determine the switching behavior of each com-
only previous results in [24] indicate that it is possible to estab- ponent. The prediction task focuses on simulating the model
lish a data-driven prognostics model to predict the RUL from forward in time via Monte Carlo simulations until the event
SF6 density data samples. under study occurs.
In this paper we focus on developing a model-based prog- In short, existing PDMP approaches for prognos-
nostics approach to integrate engineering knowledge within tics [26], [29] assume a static stochastic failure behavior and
the prediction model as in [25], instead of using data-driven do not integrate the uncertainty that may surround the failure
prognostics prediction approaches (e.g., [6]). Accordingly, we threshold specification. On the other hand, the reliability
analyze the deterioration of CBs based on I 2 t CD which community has been using PDMPs to evaluate the system
requires considering simultaneously stochastic and determin- failure probability by dynamically updating the stochastic
istic operations. failure probability of systems with deterministic operational
signals, but the system failure probability has not then been
used to make prognostic predictions.
B. Hybrid Prognostics Approaches In this paper, we seek to improve prognostics predictions
While a relatively new area, hybrid approaches have by combining dynamic reliability and uncertainty approaches
recently gained interest in the engineering prognostics arena with hybrid prognostics models. We focus on the use of
(e.g., [26], [27]). Hybrid prognostics techniques integrate the PDMPs for failure prognostics of hybrid systems because it
continuous and discrete behavior of systems, reflecting a provides the flexibility to specify the aging behavior of systems
more realistic behavior of many industrial systems whose with any PDF, and dynamic reliability approaches can be
continuous-time dynamics are affected by periodic discrete integrated in a straightforward manner.
events.
PDMPs provide a generic mathematical framework to model
hybrid systems [28]. PDMPs represent deterministic trajecto-
ries interspersed with random jumps. The deterministic motion III. I NTRODUCING N EW H YBRID
models the continuous phenomena (e.g., crack growth [26]) P ROGNOSTICS A PPROACH
while the stochastic motion models shocks which influence This section introduces PDMP, and shows how to
the deterministic part of the system at random discrete times. combine prognostics, dynamic reliability techniques, and
PDMPs have proven to be valid to create hybrid prognos- uncertainty in the failure threshold within a PDMP framework.
tics applications [26], [29]. These approaches assume that the Sections III-A and III-B review the existing literature
system failure behavior is modeled with a shock model which on PDMPs and their use for prognostics, respectively.
is designed a priori. The shock models determine the future Section III-C introduces the extensions we propose for
behavior of the system with discrete impacts on the determin- hybrid prognostics. After this, the following section applies
istic motion and it is expressed with homogeneous Poisson this framework specifically to prognostics of CB aging due
processes with constant intensity. The approach in [29] adds to CD.

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
AIZPURUA et al.: MODEL-BASED HYBRID APPROACH 1641

A. Preliminaries on PDMP Assuming that observations {y1 , . . . , yn } are available, the

A PDMP process Z(t) couples a deterministic motion likelihood of the RUL prediction is defined as
denoted with a function  and a jump process (Xn , Tn )n≥0  
L RU Ltp |Y1 = y1 , . . . , Yn = yn . (3)
where Xn represents the state of the system after the jump n
at time Tn [28]. A possible solution of (3) involves a two-step methodol-
As defined in [26], the jump process follows a Markov ogy [26]. First, it is necessary to compute the conditional
renewal process, where the next jump (Xn+1 , Tn+1 − Tn ) distribution of Zt for t > tp considering available observations
depends on the past only through the value of the last position {Y1 = y1 , . . . , Yn = yn } denoted as μy1 ,...,yn (t) and defined as
  
Xn and the process restarts at each renewal time Ti . This law μy1 ,...,yn (t) = L Zt Y1 = y1 , . . . , Yn = yn . (4)
is given by N(Xn , ·, ·), where N = (N(x, dz, dt)) is called the
renewal Markov kernel of the process [38]. Second, the reliability of the system is calculated when the
Let us consider a probability space (E, ε) which represents initial state of the system is z, denoted Rz (t), and defined as
the values of the possible states of the system. As defined the probability of the process Z being in a nonfailed state
in [26], a process Z with values in E is a PDMP if it can be
Rz (t) = P(Z(s) ∈ U ∀s ≤ t). (5)
written as follows:
Finally, the RUL likelihood given observations up to the
Z(t) = (Xn , t − Tn ), Tn ≤ t < Tn+1 (1) prediction time instant tp can be defined as follows:
 
with the following assumptions. P RU Ltp > s|Y1 = y1 , . . . , Yn = yn
1) (x, t+s) = ((x, t), s), for all (s, t) and s → (x, s) 
is right continuous with left hand limits ∀x. = Rz (s)μy1 ,...,yn (tp )dz. (6)
E
2) (Xn , Tn )n≥0 is a Markov renewal process, with T0 =
0 by convention, and with kernel N(x, dz, dt) = Therefore, the computation of the RUL distribution focuses
dFx (t)Q((x, t), dz) such as: on calculating the conditional distribution in (4) and quan-
a) dFx is the probability function of min(Sx , α(x)) tifying the reliability in (5). When considering the different
with: future conditions, only the second step is needed to recompute
i) Sx random variable with hazard rate b((x, t)); the RUL.
ii) α(x) ∈ R+ deterministic time such as
α((x, u)) = α(x) − u; C. Proposed Approach
b) Q is a probability of transition on E × E. For PDMP-based hybrid prognostics models, the reliabil-
The function dFx denotes the law of time before the next ity model in (5) has been traditionally defined using a priori
jump from position x, Q(z, ·) represents the law of the position defined shock model [26], [29]. Dynamic reliability techniques
after a jump from position z, and s denotes a time increment account for dynamic operation conditions and improve the
in the process state space. Full detail is given in [26]. accuracy of the reliability model (5) for time-varying oper-
ation conditions. Additionally, the specification of the failure
B. PDMP-Based Prognostics state U is surrounded by uncertainty criteria. This threshold
determines the final RUL value [see (2)], but to the best of
Adhering to the formal notation in [26], the system degrada-
authors’ knowledge this uncertainty has not been captured for
tion is modeled using a PDMP process denoted Z = (Zt )t∈R+
hybrid prognostics modeling approaches.
with values in (E, ε) where the random variable Zt represents
Fig. 4 shows the proposed approach for the prognostics of
the state of the system at time t.
hybrid systems, integrating the dynamic reliability, and uncer-
Prognostics involves the prediction of the RUL. If U denotes
tainty criteria with the prognostics method from the literature.
the nonfailed state of the system, then RUL at the prediction
The next sections explain each module in detail.
instant (tp ), RU Ltp , can be defined as the minimum time that
1) Prognostics: The prognostics module implements the
the PDMP process Z needs to reach the failed state in the
model in (6) taking as input the likelihood of the actual system
process state space [26]. Formally
state (diagnosis), the uncertain failure threshold, and the future
 
RU Ltp = inf s ≥ tp , Zs ∈
/ U − tp . (2) system behavior determined by the dynamic reliability and
system model blocks.
The renewal process models damage events that occur The prognostics module not only predicts the distribution of
repeatedly over time, and which further degrade the system the RUL of the asset under study (see Fig. 4), but it can also
until it reaches the failure state U . Renewals take place each predict the evolution of a property of interest before reaching
time a damage event occurs and they accumulate damage on the failure threshold, e.g., remaining number of operations of
the system [38]. However, note that there is no asset repair a CB before reaching the failure threshold.
until reaching the failure state U . 2) Diagnostics: As the system under study deteriorates it
When using PDMP for prognostics, it is necessary to will go through several health states. The diagnostics module
calculate the likelihood of the RUL prediction in (2) accord- focuses on the current health state estimation given all the
ing to available observations. At discrete observation times evidence up to the estimation time instant tp .
0 < t1 < . . . < tp partial information of the system state For industrial systems, it is very difficult to physically assess
is considered modeled by random variables {Y1 , . . . , Yn }. the state of the system. Normally, this is diagnosed through the

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
1642 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 48, NO. 9, SEPTEMBER 2018

Fig. 4. Prognostics under uncertainty and dynamic reliability driven future evolution.

collected signals and knowledge of the system deterioration aging process is defined with the differential equation
process. This process is not deterministic because there are
dL
noise terms involved in the data gathering process. Therefore = (t). (8)
the health state is represented with a PDF. As indicated in (4): dt
the system health state μy1 ,...,yn (t) is estimated evaluating the Accordingly, we can redefine the variable failure rate and
likelihood of the collected signals {y1 , . . . , yn } with respect to the PDF of the Weibull distribution changing the calendar time
the system state {Y1 , . . . , Yn } given the degradation process Zt . variable t in (7) with the aging variable L(t) [31]
This module is commonly implemented using filtering tech-  
niques such as the Bayesian particle filter [2], [5], but it can β L(t) β−1 −L(t) β
f (L(t)) = · ·e η (9)
be also implemented using other probabilistic state-estimation η η
techniques such as hidden Markov models [39] or dynamic where L(t) is determined by the solution of (8).
Bayesian networks [40]. The analytic solution of these systems is not trivial due
3) System Model and Dynamic Reliability: The system to the nonlinearities involved. A tradeoff decision is needed
model defines the degradation behavior of the asset under between the complexity of the analytic solution and com-
study. This model includes deterministic and stochastic equa- putational cost of simulations. In this paper we focus on
tions to define the behavior of the asset and inter-relationships simulation techniques based on the following basic reliability
between these equations, i.e., how shocks affect the determin- theory concepts [41].
istic motion and vice-versa (see Section IV for an example). The probability of a continuous random variable X to take
The dynamic reliability model predicts the random shock a value within [a, b] can be defined with the integration of the
instants which impact the system model of the asset under PDF fx (x)
study. Dynamic reliability approaches focus on updating  b
dynamically the PDF representing the system failure state
P[a ≤ X ≤ b] = fx (x)dx. (10)
according to operational conditions (e.g., [30]–[33]). For a
instance, the Weibull distribution allows the specification of
An alternative definition is the following:
a time-varying failure rate of the system (λ(t)) with the
following density function:  t+ t
P[t ≤ X ≤ t + t] = fx (x)dx = fx (t) t. (11)
t
  β−1 β
β t −t That is, if t is an infinitely small period, the probability
f (t) = · ·e η
(7)
η η that X is within [t, t + t] is fx (t) t.
In order to simulate the nonlinear failure PDF which is
dependent on nonlinear equations, the approximation fx (t) t
where β is the shape parameter and η is the scale parameter. can be used to evaluate the failure probability and determine
It is possible to generalize this relationship with an aging the failure occurrence time. Implementing stochastic Monte
variable L(t) that accounts for the real utilization of the asset Carlo simulations, at each simulation step t this condition is
under study [31]. In this paper, we extend this relationship evaluated as in the inverse transform sampling method [31]
for a general stress signal (t) (e.g., the current a CB must
interrupt) to be applied in prognostics studies [see (5)]. The fx (t) t ≥ r (12)

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
AIZPURUA et al.: MODEL-BASED HYBRID APPROACH 1643

where r is a random number drawn from the uniform distri- In order to quantify the maintenance benefits a cost-benefit
bution r ∼ U([0, 1]). assessment is needed because the condition-based technologies
The time instant at which the condition in (12) is satis- and engineering efforts also incur costs. The cost assess-
fied represents the time at which the stochastic jump process ment is beyond the scope of this paper. Please see [44] for
occurs, and it is denoted tjump . By the law of large num- the cost-benefit assessment of different maintenance strategies
bers [42], tjump identifies PDF occurrence time and enables including the prognostics-updated predictive maintenance.
the approximation of the nonlinear failure PDF such as in (9).
4) Failure Threshold Uncertainty: Prognostics applications IV. A PPLICATION TO C IRCUIT
are surrounded by uncertain information criteria that need to B REAKER P ROGNOSTICS
be integrated and propagated in the prediction model [43].
The CD of a CB is defined by the current in the arc between
One specific example of interest is the definition of the fail-
the fixed and moving contacts which forms while the breaker
ure zone. The exact specification of this threshold value is a
is opening (see Fig. 2). Formally, it is defined as [11]
complex task.  topen
The PDF of the failure threshold (fU (t)) indicates the
CD = i2 · dt (13)
uncertainty surrounding the failure threshold value. We have ttrip
extended the calculation of RUL in (6) to include uncer-
where i is the instantaneous current and t is the arc time. Since
tainty criteria by adding a process to evaluate and classify
arc time is difficult to measure, it can be approximated by the
predictions with randomly sampled failure threshold values
time between the CB receiving the signal to open (ttrip ) and
as follows.
the time the breaker reports being open (topen ).
1) Random Sampling: Sample a failure threshold rU ran-
The current that flows in the CB during its opening is differ-
domly from the failure threshold PDF fU (t).
ent depending on the reason for its operation. Namely, current
2) RUL Estimation: Perform RUL computations with rU .
during switching operation is generally much lower than fault
3) Classification: Take the resultant PDF of the RUL esti-
current. However, the effect of all operations is measured with
mation and classify the frequency of occurrence of the
the same three variables: 1) shock instant; 2) shock strength;
maximum RUL estimation.
and 3) shock duration.
The repeated execution of steps 1–3 generates a PDF
Network switching is the periodic operation of the CB in
(histogram) with the maximum failure probability RUL esti-
order to reconfigure the network. Its instantaneous effect on
mations. The classification of the results in step 3 can be done
the CD is small because the current (shock strength) is small
in different ways. For instance, it is possible to create a 3-D
and can be cleared quickly (shock duration). However, in the
plot which shows the different RUL PDF estimations for the
long run, reconfiguration in response to a daily load cycle can
different failure threshold values rU . However, as a more infor-
lead to a considerable number of network switching operations
mative (and intuitive) representation, RUL prediction values
that contribute to the aging of the CB. Hence, switching can
with maximum probability have been used to create another
be said to affect the CD with periodic shock instants and a
distribution function (see Section V).
constant shock strength
5) Benefits of the Approach: The proposed approach pro-
vides a more accurate estimate of the RUL of the asset iRMS if t = [tswitch , tswitch + tend_switch ]
under study (see the next section for numerical examples). i(t) = (14)
0 otherwise
Potentially this estimation may have benefits for the mainte-
nance of the asset under study through the implementation where iRMS is the root mean square current at the time of the
of fit-for-purpose maintenance planning strategies which can operation, and tswitch and tend_switch are periodic on/off times
reduce maintenance costs by operating assets for longer with that determine the shock duration.
the assurance that it will not fail [44]. Fault clearing is a stochastic event that occurs when the
As shown in the next section with a CB case study, tradi- electrical network protected by the CB experiences a fault.
tional approaches with periodical operation-based maintenance After the fault occurs, it is assumed that the circuit is returned
schedule [18] or approaches which use static and determin- to service with a constant exponential repair rate with parame-
istic operation logic [26] estimate a conservative number of ter μrep . The shock instant depends on the time of the network
remaining number of operations. Interestingly, the proposed fault, which is specified with a PDF which depends on a stress
approach updates this estimate with real usage scenarios, and signal. The stress signal (t) modifies an a priori specified
consequently it makes possible to adapt conservative calcula- PDF of circuit reliability according to the network load. In this
tions with more realistic figures improving the accuracy of the case, the chance of the circuit faulting is assumed to increase
RUL estimation. with load, but this property does not hold universally. For a
Depending on the real usage, the difference between rural network with a nonload-related reliability, for example,
dynamic reliability-based prognostics predictions and tradi- the circuit reliability PDF should be specified appropriately.
tional or static approaches varies. In the analyzed case study, it In order to calculate the future shock instants, we
is possible to run the asset under study longer with the assur- solve (7)–(11) for each simulation step t taking into account
ance that it will not fail, and undertake maintenance actions possible future stress profiles (daily load profiles) and checking
less frequently. However, it may also be the case that the real if the acceptance condition in (12) is satisfied. If the accep-
usage is more stressed than traditional or static approaches, tance test is passed, then we use the current time instant as
and this would lead to an earlier maintenance action. the occurrence of a network fault denoted as tjump .

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
1644 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 48, NO. 9, SEPTEMBER 2018

The definition of a deterministic function which defines the

fault current is not feasible for generically capturing the wide
range of types of fault which may occur. However, based
on engineering knowledge, we can define a current interval
[iA , iB ] for a given circuit and sample randomly the fault cur-
rent ifault according to ifault ∼ U([iA , iB ]). This will determine
the shock intensity of the fault clearing operation.
The open time of the CB for fault clearing can be defined
with a random variable tclear which obeys the exponential dis-
tribution tclear ∼ exp(λ). It may be expected that repeated fault
operation places particular stress on the CB, with a consequen-
tial effect on opening times. This can be modeled by (15),
which takes account of the cumulative effect of fault current
stress

d=D 
 
shock =  tjump k1 (15)
d=1

where D is the total number of network fault shocks, k1

is a normalizing constant, and (tjump ) is the stress signal
evaluated at the shock instant tjump .
These random shocks have an effect on the CB operation
by increasing the shock duration and accordingly affect-
ing the CD. Therefore, we define the shock duration as
shock duration = tclear +delay, where delay includes the effect
of cumulative shocks on the open time and is defined as
follows:

d=D 
delay = shock · e shock
k2 (16)
d=1

where shock variable is computed in (15), k2 is a normalizing

constant, and D is defined as in (15). For each shock instant,
an extra delay is added to the open time of the CB.
Fig. 5. CB prognostics algorithm.
Depending on the process under study, it is possible to
tune the normalizing constants from historical delay data so
as to match the effect of cumulative shock processes or they change the time-dependent variables in (2) with number of
can be identified from real values (e.g., spring constant in a operations and accordingly estimate the number of operations
valve [26]). A possible extension of the shock model could be until reaching the failure threshold. The proposed approach
to analyze the effect of the number of CB operations on the involves the following steps.
failure PDF in (9) (see Section VI). 1) Randomly sample the failure threshold (rU ) according
The stochastic events in (15) and (16) affect the determin- to the PDF of the failure zone fU (t).
istic motion in (13) at discrete time instants by modifying the 2) Randomly sample the initial health state (i2t_init)
shock intensity and duration. These equations are embedded according to the PDF of the actual health state
in Fig. 4 as follows. μy1 ,...,yn (t).
1) The system model block embeds (13)–(16). 3) Check the network switching condition and if satisfied,
2) The dynamic reliability formulation is given by (7)–(9) increase the CD and CB operation counter.
and the computation of the circuit failure time (tjump ) is 4) Check if a network fault has occurred.
calculated with the condition in (12). The repair behavior a) Obtain the jump time tjump with function
of the circuit is defined with an exponential distribution GetJumpTime and check if tjump > 0. GetJumpTime
with a constant repair rate μrep . implements the condition in (12), i.e., tjump > 0 iff
3) Diagnostics and uncertainty are expressed with PDFs the condition in (12) is true. If true:
μy1 ,...,yn (t) and fU (t), respectively. i) calculate the stochastic open time for fault
Fig. 5 outlines the flowchart of the CB prognostics algorithm clearing (tclear ) with function GetOpenTime
making use of the proposed approach. using inversion sampling [42]: tclear = F −1 (r),
Apart from the RUL estimation, we also calculate the where F −1 is the quantile function and r ∼
remaining number of CB operations. To this end, we only U([0, 1]);

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
AIZPURUA et al.: MODEL-BASED HYBRID APPROACH 1645

TABLE I
VARIABLES FOR CB P ROGNOSTICS E XPERIMENTS

Fig. 6. I 2 t state at prediction time instant tp = 9.4 years.

ii) for each shock occurrence (identified by the

sshock variable) calculate its effect on the delay,
increase the CD accordingly, and increase the
CB operation counter.
5) RUL calculation for the kth Monte Carlo trial [see (2)].
6) RUL and remaining CB operations probability calcula-
tion after K Monte Carlo trials.
7) Classification of uncertainty results first creating a PDF Fig. 7. Analyzed daily load profiles.
with kernel density estimation [45] and then taking
the RUL and remaining CB operation values with the
maximum probability.
8) Estimate the PDF of the RUL and remaining number of
CB operations with randomly sampled failure threshold
values.

Fig. 8. Effect of network faults on the open time.

V. C ASE S TUDY
The Power Networks Demonstration Centre (PNDC) is an
11-kV/400-V test facility for demonstrating smart grid tech- i.e., traditional PDMP-based prognostics prediction method
nologies, located near Glasgow, U.K. It was designed to (e.g., [26]). In contrast, dynamic reliability profiles modify
accommodate significant levels of automation and communi- the static behavior according to the applied load. Fig. 7 shows
cations, embedded generation, and fault-throwing capability, the different types of daily load profiles [(t) in (8)].
with the intention of trialling equipment and procedures for Different current load profiles impact the PDMP model in
a distribution network of the future [46]. One of the 11-kV different ways and accordingly determine the future RUL and
CBs at PNDC was selected for testing the hybrid prognos- remaining number of CB operation values.
tics approach. After conducting a workshop with engineers, Fig. 8 shows the effect of network fault events on the open
Table I displays the variables for CB prognostics experiments time delay of the CB. Note that the delay increases only with
and their values. network faults and not with switching operations.
According to the fault levels experienced at PNDC (Table I, In Fig. 8, the CB switching duration starts to deviate from
ifault ) and manufacturer specifications (Fig. 3) the standard normal behavior after 450 further network fault events approxi-
maintenance strategy for the chosen breaker is to replace it mately (see low dashed line). After this threshold value, the CB
every 10 years or after performing 10 000 operations. starts to operate in a degraded mode owing to the cumulated
In an online context, the actual health of the CB can be network faults. These faults increase the delay exponentially
monitored through I 2 t data samples {y1 , . . . , yp }, and predic- [see (16)] until the CB reaches the failure limit of the CD
tions of RUL made on an on-going basis. For the purpose of (see upper dashed line, CB Fault event). Due to lack of
this paper, we assume a prediction time instant where the CB real information, we have not taken into account previous
is diagnosed at 94% of the preplanned maintenance period. cumulative delay. However, if the designer has this informa-
That is, the remaining number of CB operations is 600 or tion, the accumulated delay should be added to the process
equivalently the RUL is 0.6 years. The health state estimation in Fig. 8.
indicated in (4) can be implemented using filtering techniques Taking a fixed failure threshold value [mean value of fU (t)
(see Section III-C2). Fig. 6 shows the estimated health state in Table I], Fig. 9 shows different predictions of the remain-
of the asset at the prediction time instant. ing number of operations according to the daily load profiles
For future prediction instants, we compare the use of both shown in Fig. 7. While the constant load profile specified
static and dynamic reliability profiles. The static profile imple- according to the static reliability model shows the most con-
ments the shock instants according to a constant load profile, servative estimate, different predictions estimated with typical

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
1646 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 48, NO. 9, SEPTEMBER 2018

Fig. 11. RUL estimation with uncertainty prediction criteria.

Fig. 9. Prediction of remaining number of operations.

RUL estimations in Fig. 10 and then we apply kernel density
estimation (see Section III-C and Fig. 5, step 8). Fig. 11 shows
the final PDF which integrates uncertainty criteria.
Fig. 11 shows that when the uncertainty in the failure thresh-
old is included, the shape and maximum likelihood values
of the PDF of RUL change accordingly. In this case, the
prediction with failure threshold uncertainty information pre-
dicts an extended RUL value (dashed line) compared with the
fixed failure threshold estimation (solid line).
The final effect of the uncertain failure threshold level on
the RUL prediction depends on the adopted failure threshold
distribution fU (t). If the designer has a better informed knowl-
edge of the failure threshold, fit-for-purpose failure threshold
Fig. 10. RUL estimations with randomly sampled failure threshold values.
distributions may be used which can lead to more accurate
results.
As confirmed with the results shown in this section, the
daily load profiles show an extended number of remaining CB combined use of dynamic reliability and hybrid prognos-
operations. This highlights the importance of accurate load tics models improves the accuracy with respect to periodic
forecasting on CB prognostics. operation-based maintenance schedules and hybrid prognos-
Note that the static reliability driven prediction in Fig. 9 tics models with static reliability models (Fig. 9). Besides, the
does not match with the preplanned maintenance strategy of integration of the uncertainty information in the failure thresh-
600 remaining CB operations. This is because the default old enables the evaluation of the effect of alternative failure
maintenance strategy implements an average estimation of CD threshold values on the final prediction estimations (Fig. 10)
levels with fixed time intervals and constant current values. and prediction of the most likelihood useful life according to
Using condition-based maintenance strategies, even with static maximum RUL estimation values (Fig. 11).
reliability models, gives less pessimistic results. The static While the models and algorithms presented in Section III
PDMP model includes random intervals for fault current and are generic, Sections IV and V show the application of the
open time affected by random shocks. When we adapt the proposed model to CBs. The particular degradation equations
PDMP model with dynamic daily load profiles, we can see and dynamic operation conditions depend on the asset under
that the number of operations increases in proportion to the study. However, the theoretical framework can be generally
network load. applied to hybrid systems which degrade according to deter-
It should also be noted that the values in Fig. 9 show the ministic and stochastic events, with the particularity that the
remaining number of operations until CB failure, while main- stochastic degradation process is modulated by deterministic
tenance strategies are required to adopt a safety margin for a signals, such as usage, load, or temperature.
timely maintenance [47]. In the case study we performed 104 Monte Carlo iterations
The experiments in Fig. 9 have been performed with a fixed (K in Fig. 5), for each of the 100 random samples of the
failure threshold. However, this assumption may not be realis- initial heath state (I in Fig. 5), and these were repeated for
tic and we include uncertainty estimations based on the PDF 25 randomly sampled failure threshold values (J in Fig. 5),
of the failure threshold, fU (t). Taking the PDMP model with which gives a total of 25 × 106 iterations.
static load as a reference model and focusing on the RUL The main limitation of the proposed approach is the sim-
prediction of the CB, Fig. 10 shows different estimations with ulation time. The lower the current load profile, the longer
randomly sampled failure threshold values [see Table I, fU (t)]. the simulation time. With a standard desktop Intel i7 with 8
Each of the random failure threshold trials in Fig. 10 cor- cores, simulation times range from 8 h (static load in Fig. 7)
responds to a randomly sampled value from fU (t). While this up to 240 h (load C in Fig. 7). There are also other factors
information is useful to see the effect of the failure threshold that affect the simulation time:
on the RUL prediction, as a more informative representa- 1) Health State at Prediction Time: The better the health
tion, we take the maximum likelihood value for each of the of the asset, the longer the simulation time.

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
AIZPURUA et al.: MODEL-BASED HYBRID APPROACH 1647

2) Failure Threshold: The greater the failure threshold the e.g., dependent on failed components or available repair
longer the simulation time. resources.
3) Accuracy of the Results: The higher the required accu- 5) Simulation Time: Speed up simulations with dynamic
racy, the higher the simulation time. stopping criteria such as in [48], where the number
Although maintenance is normally planned months ahead of of Monte Carlo iterations are determined dynamically
the failure occurrence, depending on the prognostics prediction based on accuracy monitoring and confidence levels.
horizon, the simulation time may be a critical factor. In order
to alleviate long simulation periods, we have used MATLAB’s ACKNOWLEDGMENT
parallel computing toolbox. However, we also plan to study
analytical techniques to speed-up the simulation time (see the The authors would like to thank the Editor and all the
next section). reviewers for their valuable comments that helped to improve
the clarity and completeness of this paper.

VI. C ONCLUSION R EFERENCES

CBs are essential for the correct behavior of the power [1] F. Li et al., “Smart transmission grid: Vision and framework,” IEEE
network. CBs exhibit a hybrid operation mode which inte- Trans. Smart Grid, vol. 1, no. 2, pp. 168–177, Sep. 2010.
grates stochastic and deterministic operations. Accordingly we [2] V. M. Catterson, J. Melone, and M. S. Garcia, “Prognostics of trans-
former paper insulation using statistical particle filtering of on-line data,”
have presented a hybrid approach for prognostics of CBs using IEEE Elect. Insul. Mag., vol. 32, no. 1, pp. 28–33, Jan./Feb. 2016.
PDMPs. [3] A. D. Kenyon, V. M. Catterson, S. D. J. McArthur, and J. Twiddle, “An
The proposed approach extends existing hybrid prognos- agent-based implementation of hidden Markov models for gas turbine
condition monitoring,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 44,
tics techniques in two ways. On the one hand, it enables a no. 2, pp. 186–195, Feb. 2014.
more accurate RUL forecast through integration with dynamic [4] G. Vachtsevanos, F. Lewis, M. Roemer, A. Hess, and B. Wu, Intelligent
reliability techniques. That is, we update a priori defined Fault Diagnosis and Prognosis for Engineering Systems. Hoboken, NJ,
fault conditions dynamically according to operation signals, USA: Wiley, 2006.
[5] M. J. Daigle and K. Goebel, “Model-based prognostics with concurrent
e.g., daily load profiles of the power network. On the other damage progression processes,” IEEE Trans. Syst., Man, Cybern., Syst.,
hand, the proposed approach evaluates the effect of uncer- vol. 43, no. 3, pp. 535–546, May 2013.
tain failure evaluation levels in the prognostics estimation [6] S.-J. Wu, N. Gebraeel, M. A. Lawley, and Y. Yih, “A neural network inte-
grated decision support system for condition-based optimal predictive
process in order to handle the uncertainty that surrounds the maintenance policy,” IEEE Trans. Syst., Man, Cybern. A, Syst., Humans,
specification of a failure threshold. vol. 37, no. 2, pp. 226–236, Mar. 2007.
The presented approach enables fit-for-purpose maintenance [7] G. Haddad, P. A. Sandborn, and M. G. Pecht, “An options approach for
decision support of systems with prognostic capabilities,” IEEE Trans.
planning for the CB, knowing that the predictions are tai- Rel., vol. 61, no. 4, pp. 872–883, Dec. 2012.
lored to the specific operational and failure circumstances of [8] K. T. Huynh, I. T. Castro, A. Barros, and C. Bérenguer, “On the use of
a given breaker. In the analyzed case study, results show that mean residual life as a condition index for condition-based maintenance
the predictions incorporating dynamic reliability and uncer- decision-making,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 44, no. 7,
pp. 877–893, Jul. 2014.
tainty management lead to longer time horizon predictions of [9] Megger. Circuit Breaker Testing Guide. Accessed on Jan. 23, 2017.
RUL, compared to predictions which do not use dynamic relia- [Online]. Available: www.megger.com
bility concepts. However, the final prediction always depends [10] K. Harker, Power System Comissioning and Maintenance Practice, Inst.
Eng. Technol., Lucknow, India, 1998.
on the specific usage of the breaker and knowledge of the [11] R. H. Kaufmann, “The magic of I2t,” IEEE Trans. Ind. Gen. Appl.,
failure threshold. This is why specific operational conditions vol. IGA-2, no. 5, pp. 384–392, Sep. 1966.
and failure threshold specification have to be included on a [12] T. E. Wierman, D. M. Rasmuson, and N. B. Stockton, “Common-cause
failure event insights—Circuit breakers, NUREG/CR-6819,” Idaho Nat.
case-by-case basis. Eng. Environ. Lab, Idaho Falls, ID, USA, Tech. Rep. NUREG/CR-6819,
The confidence depends in part on the accuracy of forecast- vol. 1, 2003.
ing the circuit load profile. While traditionally the load profile [13] G. D. Camps, “The development of a methodology to determina the
maintenance strategy for high voltage circuit breakers,” Ph.D. dis-
of a given circuit may remain relatively static over time, tech- sertation, Univ. Strathclyde, Inst. Energy Environ., Glasgow, U.K.,
nologies associated with smart grids such as electric vehicles 2010.
and demand-side management mean that loading is varying [14] T. M. Lindquist, L. Bertling, and R. Eriksson, “Circuit breaker failure
more than ever before. However, this presents another reason data and reliability modelling,” IET Gener. Transm. Distrib., vol. 2, no. 6,
pp. 813–820, Nov. 2008.
for tracking actual CB wear instead of planning maintenance [15] P. Choonhapran and G. Balzer, “Availability of HV circuit-breakers: The
on a periodic schedule. application of Markov model,” in Proc. IEEE Power Eng. Soc. Gen.
Future work can address the following points. Meeting, Tampa, FL, USA, 2007, pp. 1–6.
[16] J. I. Aizpurua and V. Catterson, “Towards a methodology for design of
1) Validation of Results: Validate predictions with real run- prognostics systems,” in Proc. Annu. Conf. Prognostics Health Manag.
to-failure CB data. Soc., vol. 6. 2015, pp. 196–209.
2) Shock Models: Analyze the effect of making the PDF of [17] P. E. Labeau, C. Smidts, and S. Swaminathan, “Dynamic reliability:
Towards an integrated platform for probabilistic risk assessment,” Rel.
failure dependent on the number of operations. Eng. Syst. Safety, vol. 68, no. 3, pp. 219–254, 2000.
3) System-Level Prognostic Prediction: Integrate other fail- [18] Schneider Electric. GenieEvo Range Metalclad, Vacuum Indoor
ure modes (e.g., SF6 density, trip coil current) to obtain Switchgear. Installation, Operation and Maintenance Instructions.
a complete prognostic model of the CB. Accessed on Jan. 23, 2017. [Online]. Available: http://goo.gl/pospcH
[19] M. Kezunovic et al., “Automated monitoring and analysis of cir-
4) Variable Repair Rates: Apply dynamic reliability tech- cuit breaker operation,” IEEE Trans. Power Del., vol. 20, no. 3,
niques to model condition-dependent repair rates, pp. 1910–1918, Jul. 2005.

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.
1648 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 48, NO. 9, SEPTEMBER 2018

[20] S. M. Strachan, S. D. J. McArthur, B. Stephen, J. R. McDonald, [43] S. Sankararaman, “Significance, interpretation, and quantification of
and A. Campbell, “Providing decision support for the condition-based uncertainty in prognostics and remaining useful life prediction,” Mech.
maintenance of circuit breakers through data mining of trip coil cur- Syst. Signal Process., vols. 52–53, pp. 228–247, Feb. 2015.
rent signatures,” IEEE Trans. Power Del., vol. 22, no. 1, pp. 178–186, [44] J. I. Aizpurua, V. M. Catterson, F. Chiacchio, and D. D’Urso,
Jan. 2007. “A cost-benefit approach for the evaluation of prognostics-updated main-
[21] P.-C. Lin, J.-C. Gu, and M.-T. Yang, “Intelligent maintenance model tenance strategies in complex dynamic systems,” in Proc. Eur. Safety Rel.
for condition assessment of circuit breakers using fuzzy set theory Conf. (ESREL), Glasgow, U.K., 2016, pp. 1064–1071.
and evidential reasoning,” IET Gener. Transm. Distrib., vol. 8, no. 7, [45] J. S. Kim and C. Scott, “Robust kernel density estimation,” in Proc.
pp. 1244–1253, Jul. 2014. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), Las Vegas,
[22] A. A. Razi-Kazemi, M. Vakilian, K. Niayesh, and M. Lehtonen, “Data NV, USA, 2008, pp. 3381–3384.
mining of online diagnosed waveforms for probabilistic condition assess- [46] C. D. Booth, F. Coffele, and G. M. Burt, “The power networks demon-
ment of SF6 circuit breakers,” IEEE Trans. Power Del., vol. 30, no. 3, stration centre: An environment for accelerated testing, demonstration
pp. 1354–1362, Jun. 2015. and validation of existing and novel protection and automation systems,”
[23] P. Westerlund, P. Hilber, T. Lindquist, and S. Kraftnät, “A review of in Proc. IET Int. Conf. Develop. Power Syst. Protect., Copenhagen,
methods for condition monitoring, surveys and statistical analyses of Denmark, 2014, pp. 1–6.
disconnectors and circuit breakers,” in Proc. Int. Conf. Probabilistic [47] J. I. Aizpurua and V. M. Catterson, “On the use of probabilistic model-
Methods Appl. Power Syst., Durham, U.K., Jul. 2014, pp. 1–6. checking for the verification of prognostics applications,” in Proc.
[24] S. E. Rudd, V. M. Catterson, S. D. J. McArthur, and C. Johnstone, IEEE Int. Conf. Intell. Comput. Inf. Syst., Cairo, Egypt, Dec. 2015,
“Circuit breaker prognostics using SF6 data,” in Proc. IEEE Power pp. 7–13.
Energy Soc. Gen. Meeting, San Diego, CA, USA, Jul. 2011, pp. 1–6. [48] I. Meedeniya, A. Aleti, and L. Grunske, “Architecture-driven reliability
optimization with uncertain model parameters,” J. Syst. Softw., vol. 85,
[25] J. Luo, K. R. Pattipati, L. Qiao, and S. Chigusa, “Model-based prognostic
no. 10, pp. 2340–2355, 2012.
techniques applied to a suspension system,” IEEE Trans. Syst., Man,
Cybern. A, Syst., Humans, vol. 38, no. 5, pp. 1156–1168, Sep. 2008.
[26] A. Lorton, M. Fouladirad, and A. Grall, “A methodology for probabilistic
model-based prognosis,” Eur. J. Oper. Res., vol. 225, no. 3, pp. 443–454, Jose Ignacio Aizpurua received the Eng., M.Sc.,
2013. and Ph.D. degrees from Mondragon University,
[27] M. Daigle, I. Roychoudhury, and A. Bregon, “Model-based prognostics Arrasate-Mondragon, Spain, in 2010, 2012, and
of hybrid systems,” in Proc. Annu. Conf. Prognostics Health Manag. 2015, respectively.
Soc., vol. 6. 2015, pp. 57–66. He is a Research Associate with the Institute for
[28] M. H. A. Davis, Markov Models and Optimization (Monographs on Energy and Environment, University of Strathclyde,
Statistics and Applied Probability). London, U.K.: Chapman & Hall, Glasgow, U.K. He was a Visiting Researcher with
1993. the Dependable Systems Research Group, University
[29] D. N. Nguyen, L. Dieulle, and A. Grall, “Remaining useful lifetime of Hull, Hull, U.K., in 2014. His current research
prognosis of controlled systems: A case of stochastically deteriorat- interests include prognostics, dependability, condi-
ing actuator,” Math. Problems Eng., vol. 2015, 2015, Art. no. 356916, tion monitoring, and systems engineering.
doi: 10.1155/2015/35691.
[30] J. Chiquet and N. Limnios, “A method to compute the transition func-
tion of a piecewise deterministic Markov process with application to Victoria M. Catterson (S’03–M’06–SM’12)
reliability,” Stat. Probab. Lett., vol. 78, no. 12, pp. 1397–1403, 2008. received the B.Eng. (Hons.) and Ph.D. degrees
[31] F. Chiacchio et al., “SHyFTA, a stochastic hybrid fault tree automa- from the University of Strathclyde, Glasgow, U.K.,
ton for the modelling and simulation of dynamic reliability problems,” in 2003 and 2007, respectively.
Expert Syst. Appl., vol. 47, pp. 42–57, Apr. 2016. She is a Senior Lecturer with the Institute for
[32] H. Chraibi, “Dynamic reliability modeling and assessment with Energy and Environment, University of Strathclyde.
PyCATSHOO: Application to a test case,” in Proc. Probabilistic Safety Her current research interests include condition
Assess. Manag. (PSAM), Tokyo, Japan, 2013, pp. 1–10. monitoring, diagnostics, and prognostics for power
[33] M. Bouissou and X. de Bossoreille, “From Modelica models to depend- engineering applications.
ability analysis,” IFAC-PapersOnLine, vol. 48, no. 7, pp. 37–43, 2015.
[34] D. Codetta-Raiteri and L. Portinale, “Approaching dynamic reliability
with predictive and diagnostic purposes by exploiting dynamic Bayesian
networks,” Proc. Inst. Mech. Eng. O J. Risk Rel., vol. 228, no. 5, Ibrahim F. Abdulhadi received the M.Eng. and
pp. 488–503, 2014. Ph.D. degrees in electronic and electrical engineer-
[35] D. Codetta-Raiteri, “Modelling and simulating a benchmark on dynamic ing from the University of Strathclyde, Glasgow,
reliability as a stochastic activity network,” in Proc. Eur. Model. Simulat. U.K., in 2007 and 2013, respectively.
Symp. (EMSS), Rome, Italy, 2011, pp. 545–554. He is currently a Research and Development
[36] D. Codetta-Raiteri and A. Bobbio, “Solving dynamic reliability problems Engineer with the Power Networks Demonstration
by means of ordinary and fluid stochastic Petri nets,” in Proc. Eur. Safety Centre, University of Strathclyde. He has experience
Rel. Conf., vol. 1. Gdańsk, Poland, 2005, pp. 381–389. researching for U.K. distribution and transmission
[37] Q. Gaudel, E. Chanthery, and P. Ribot, “Hybrid particle Petri nets for network operators. His current research interests
systems health monitoring under uncertainty,” Int. J. Prognostics Health include power system protection, real-time power
Manag., vol. 6, no. 22, p. 20, 2015. system simulation, hardware in the loop testing, and
[38] S. Ross, Introduction to Probability Models. London, U.K.: Academic communications applications in smart grids.
Press, 2006.
[39] F. Camci and R. B. Chinnam, “Health-state estimation and prognostics
in machining processes,” IEEE Trans. Autom. Sci. Eng., vol. 7, no. 3, Maria Segovia Garcia received the Ph.D. degree
pp. 581–597, Jul. 2010. from the University of Granada, Granada, Spain,
[40] D. Codetta-Raiteri and L. Portinale, “Dynamic Bayesian networks for in 2009.
fault detection, identification, and recovery in autonomous spacecraft,” She is a Knowledge Exchange Associate with
IEEE Trans. Syst., Man, Cybern., Syst., vol. 45, no. 1, pp. 13–24, the Power Networks Demonstration Centre (PNDC),
Jan. 2015. University of Strathclyde, Glasgow, U.K. She was
[41] M. Rausand and A. Høyland, System Reliability Theory: Models, a Post-Doctoral Research Fellow with the Nuclear
Statistical Methods and Applications, 2nd ed. Hoboken, NJ, USA: Wiley, Metrology Department, Free University of Brussels,
2003. City of Brussels, Belgium, before joining the PNDC
[42] E. Zio, The Monte Carlo Simulation Method for System Reliability in 2013. Her current research interests include
and Risk Analysis (Series in Reliability Engineering). London, U.K.: modeling assets deterioration, estimation of main-
Springer, 2013. tenance actions impact, and data analysis.

Authorized licensed use limited to: Northeastern University. Downloaded on April 02,2024 at 01:35:29 UTC from IEEE Xplore. Restrictions apply.