Professional Documents
Culture Documents
Abstract—Prognostics focuses on predicting the future perfor- SVM Support Vector Machine
mance of a system, specifically the time at which the system no
long performs its desired functionality, its time to failure. As an SOC State of Charge
important aspect of prognostics, remaining useful life (RUL) pre- SOH State of Health
diction estimates the remaining usable life of a system, which is
essential for maintenance decision making and contingency miti- EKF Extended Kalman Filter
gation. A significant amount of research has been reported in the
literature to develop prognostics models that are able to predict BP Back Propagation
a system’s RUL. These models can be broadly categorized into TLFN Time-Lagged Feedforward Network
experience-based models, date-driven models, and physics-based
models. However, due to system complexity, data availability, and FIS Fuzzy Inference System
application constraints, there is no universally accepted best model MQE Minimum Quantization Error
to estimate RUL. The review part of this paper specifically focused
on the development of hybrid prognostics approaches, attempting PDF, pdf Probability Density Function
to leverage the advantages of combining the prognostics models in
the aforementioned different categories for RUL prediction. The MD Mahalanobis Distance
hybrid approaches reported in the literature were systematically RVM Relevance Vector Machine
classified by the combination and interfaces of various types of
prognostics models. In the case study part, a hybrid prognostics RBPF RaoBlackwellized Particle Filter
method was proposed and applied to a battery degradation case to ANFIS Adaptive Neuro Fuzzy Inference System
show the potential benefit of the hybrid prognostics approach.
SVR Support Vector Regression
Index Terms—Condition monitoring, hybrid approach, predic-
tive maintenance, prognostics, remaining useful life prediction. EOL End of Life
MAE Mean Absolute Error
Manuscript received May 27, 2013; revised July 19, 2013; accepted
Predicted measurements
September 10, 2013. Date of publication January 15, 2014; date of current
version February 27, 2014. Associate Editor: L.Cui.
Prediction step
L. Liao is with the Palo Alto Research Center (PARC, a Xerox Company) Probability density function
(e-mail: linxia.liao@parc.com; liaolinxia@gmail.com).
F. Köttig is with the Dresden University of Technology (TUD), Germany, and Warburg impedance
is also with the Siemens Corporation, Corporate Technology U.S.
Color versions of one or more of the figures in this paper are available online Charge transfer resistance
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TR.2014.2299152
Electrolyte resistance
0018-9529 © 2014 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: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
192 IEEE TRANSACTIONS ON RELIABILITY, VOL. 63, NO. 1, MARCH 2014
B. Data-Driven Model
I. INTRODUCTION TO PROGNOSTICS APPROACH
Data-driven models rely only on previously observed data to
P ROGNOSTICS technology covers many aspects such as predict the projection of a system’s state or to match similar
yielding advanced warning of impending failures, and es- patterns in the history to infer RUL. Data-driven models include
timating remaining useful life, which ultimately result in in- but are not limited to statistical models, reliability functions, and
creased availability, reliability, and safety; and reduced main- artificial intelligence models.
tenance and logistics cost. As mentioned in ISO13381–1, prog- Statistical models such as Hidden Markov Models (HMM)
nostics is “an estimation of time to failure and risk for one or have been applied to prognostics problems. Medjaher et al.
more existing and future failure modes” [1]. Remaining useful [5] presented a data-driven prognostics method for remaining
life (RUL) was defined as “the length from the current time useful life prediction with applications to bearings. Mixtures of
to the end of the useful life” in [2]. As a very important part Gaussian hidden Markov models (MoG-HMMs), represented
of prognostics, RUL prediction has been applied to many sys- by dynamic Bayesian networks (DBN), were used to assess the
tems such as military and aerospace systems, manufacturing bearing health state based on time domain or frequency domain
equipment, power systems, and electronics. In general, RUL features extracted from vibration measurements. A set of five
prediction models in the literature can be categorized into ex- run-to-failure accelerated life test datasets was used for training
perience-based models, date-driven models, and physics-based of the MoG-HMMs, i.e., for estimation of the behavior model
models, as shown in Fig. 1. parameters corresponding to different health states (offline
training). The prognostics performance was evaluated with a
sixth dataset. Using the learned models, an appropriate model
A. Experience-Based Model
was selected based on the likelihood of the models given the
Experience-based models correlate expert knowledge and observations, and the remaining useful life was estimated by
engineering experience (often documented as IF-THEN rules summarizing the stay durations in each step (until failure)
directly from domain experts) with the observed situation to of the selected DBN. Medjaher et al. chose MoG-HMMs to
infer RUL from historical measurements or events. Expert overcome the limitations of traditional HMMs only capable of
systems and fuzzy logic methods can be considered as experi- handling discrete observations. Peng and Dong [6] presented
ence-based models because domain knowledge is relied on for a prognostics approach of hidden Markov and grey models
both methods. to predict the health states of pumps. The traditional hidden
Biagetti and Sciubba [3] presented an expert system incor- Markov state transition model was extended by an aging factor
porating a knowledge base, which is able to detect both faults to consider the deterioration of the device. This aging factor
and insidious degradation, i.e., to perform both diagnostics and was iteratively optimized with a statistical algorithm. The RUL
prognostics for an application in plants. IF-THEN rules were was estimated by analyzing the expected durations when the
defined with estimated fault indicators, and Fuzzy representa- device stayed in the healthy states until a failure occurred.
tion was used for fault diagnosis. Prognostics was performed Features were extracted from vibration signals using wavelet
by projecting the future state based on the present state and analysis, and the wavelet coefficients were used as the inputs
foreseen operative conditions. However, RUL was not deduced to the HMMs. These models can model different health stages
from the proposed method. The rules defined in expert systems (e.g., normal, crack initialization, growth, and failure) without
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
LIAO AND KÖTTIG: REVIEW OF HYBRID PROGNOSTICS APPROACHES FOR REMAINING USEFUL LIFE PREDICTION OF ENGINEERED SYSTEMS 193
the assumption of monotonic degradation. The specific fault suming. The reliability function may require assumptions such
progression in each stage is not required, but it requires data as the times to failure must be –independent and identical dis-
collected from different stages for training. A large amount tributed, all covariates that affect the hazard rate must be in-
of data, which is proportional to the different health stages, cluded in the model, and the effect of covariates is statistically
is required to train the model. The unforeseen failure modes time dependent.
cannot be modeled. Artificial intelligent methods such as neural networks were
Statistical models can further utilize time-series regression presented by Gebraeel et al. [13] to predict the RUL of bear-
techniques. Liu et al. [7] proposed an advanced autoregressive ings. Heng et al. [14] presented a prognostics approach using
time series model with a nonlinear accelerated degradation feedforward neural networks on pump vibration data. The
factor for remaining useful life estimation of lithium-ion bat- model incorporated population characteristics and suspended
teries. The degradation factor was obtained with curve fitting, condition trending data of historical units into prognostics.
and added to the AR model output to extend the linear AR The feedforward neural network was trained to estimate the
model, and match the battery degradation process. Yan et al. future survival probabilities, and RUL was calculated until the
[8] presented a method of using an auto-regressive moving predicted survival probability was less than 0.5. He et al. [15]
average (ARMA) model to predict the RUL of an elevator door. presented a method for health monitoring and prognostics of an
An ARMA model was used to predict the trend of the failure axial piston pump, combining empirical mode decomposition
probability, which was modeled from sensor data using logistic (EMD) of time series, particle swarm optimization (PSO), and
regression, until the trend of the failure probability reached a support vector machines (SVMs). Using the oil return flow as
predefined failure threshold. Lamoureux et al. [9] defined a an indicator of the wear condition, EMD was used to assess
physics-based health indicator for prognostics of an aircraft the pump’s health state, and system health state trending was
engine fuel pumping unit by applying a linear regression be- accomplished by SVM regression based on the intrinsic mode
tween a system degradation severity indicator and the measured functions decomposed by the EMD. He et al. used PSO to
health indicator (in this case, pump rotation speed). Typical obtain optimal parameters for the SVM radial basis function
ARMA models are effective for short-time prediction, while kernel. Artificial intelligent methods have been extensively
they are not reliable for long-term prediction. The predicted studied in the literature, which reveals their advantages and
value is assumed to be a linear function of previous observa- disadvantages. For example, neural networks can be used to
tions, with random errors. The stationary data assumption and model non-linear system functions, and are widely applied
the auto-correlation –independence assumption need to be to engineered systems. The feedback layer within an Elman
verified. The method of linear regression is easy to implement, Recurrent Neural Network enables it to model a non-stationary
and the prediction result is easy to explain. However, at the process. It is an effective tool to model complex systems which
fault initialization stage, the health indicator is typically noisy involve non-linear behavior and unstable processes, while cur-
and non-monotonic. Hence, a linear regression may not fit in rently there is no standard method to determine the structure of
this situation. Changes in the operating conditions (e.g., speed, the network, and it requires sufficient computational resources.
loading) and changes in the failure modes may affect the fault SVM is capable of finding an optimized hyper-plane in the
propagation, which makes the prediction of linear regression projected space that can separate the target points, which can
unreliable. only be classified by a non-linear boundary in the original
Reliability functions such as Weibull distributions have also feature space. Currently, there is no standard method to choose
been used for data-driven prognostics. Guo et al. [10] presented the kernel function for SVM, which is the key procedure of a
a three-parameter Weibull failure rate function for wind turbine successful SVM application.
reliability assessment. The proposed model dealt better with in- Furthermore, similarity based pattern matching methods have
complete field failure data than the traditional Weibull model, also been proposed for RUL prediction when historical run-to-
which in fact is a special case of the introduced three-param- failure data are abundant, such as the match matrix method pre-
eter model. Sutherland et al. [11] presented an approach for sented by Liu et al. [16], and a similarity based method pro-
deriving and using accurate Weibull distributions by means of posed by Wang et al. [17]. The key point of those methods
identification of subpopulations of related systems in the main- was to match the degradation pattern, which was represented
tenance database for condition-based maintenance of motors. by a signal or modeled health indicator, to the historical run-to-
Goode et al. [12] presented a RUL prediction method using failure datasets. The best matching pattern was found to derive
a statistical model for an application of pumps in a hot strip RUL, or weighted multiple patterns were used to calculate the
steel mill. A Weibull distribution was used to model the time-to- RUL. The similarity based prediction method is fundamentally
failure. RUL was then calculated during the potential failure to a pattern matching method. When there are a large amount of
functional failure interval, combining vibration data with reli- run-to-failure datasets from various operating conditions and
ability data. The reliability functions, which require data asso- failure modes, it can generate very accurate predictions. How-
ciated with individual failure mode(s), can be developed with ever, run-to-failure data are prohibitive and impractical in many
sufficient historical data. The method incorporates system age cases. The final prediction result also largely depends on the nor-
and –independent hazards into RUL prediction. It is easy to malization process, which calculates the average RUL from all
build a statistically adequate model; however, the model may matched historical patterns.
not follow the underlying degradation process. The parame- Data-driven models are derived only from historical data,
ters are often selected by experts, and that process is time con- and are applicable when data are sufficiently abundant. Similar
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
194 IEEE TRANSACTIONS ON RELIABILITY, VOL. 63, NO. 1, MARCH 2014
models can be applied to other systems without understanding sampling scheme for particle filter based prognostic approaches,
the complex physics. The output generated from real monitoring namely using the unscented transform, Daigle and Goebel also
data tends to give more precise information than those obtained presented a detailed model of the solenoid valve and extended
from experience or maintenance feedback. However, most re- it according to the damage evolving during the valve’s lifetime,
sults of data-driven models are not easy to explain or to be re- as presented in [22]. The measurement models, which estab-
lated to any physical meaning. In prognostics applications, there lish the mapping from measurement to the internal system state,
are also challenges in determining thresholds, deciding data to were the system states themselves, plus measurement noise. In
describe normal behavior, and solving over-fitting issues. the literature, data-driven models have been widely used to infer
the measurement models, which is addressed in Section II-D.1.
C. Physics-Based Model Physics-based models are derived directly from first prin-
Physics-based models are approaches that involve the knowl- ciples and an understanding of physical mechanisms. When
edge of a system’s failure mechanisms (e.g., crack growth) to available and sufficiently complete, behavioral models tend
build a mathematical description of the system’s degradation to significantly outperform other types of models [23]. Addi-
process to estimate the RUL. The mathematical models quan- tionally, changes in model output as described by the residuals
titatively characterize a system’s behavior using physics or first tend to have a direct (or easily translatable) physical meaning
principles. The identification of the model parameters usually [24]. However, physics-based models may not be available to
requires specifically designed experiments and extensive empir- complex systems due to the lack of understanding of all failure
ical data. In prognostics, condition data are often used to iden- modes and behaviors under a range of operating conditions.
tify and update the model parameters using statistical methods Model parameter identification also requires extensive experi-
(e.g., regression, Bayesian update). In our research, prognostics ment. In addition, a physics-based model is often built case by
models which involve first principles and statistical methods for case. Hence, it is not generally applicable to a different system
parameter identification are still categorized as physics-based without a significant amount of effort.
models.
Zhao et al. [18] presented a prognostics method for remaining II. HYBRID PROGNOSTICS APPROACH
useful life prediction for gears. The crack growth process was
simulated to yield normally distributed crack lengths, which More literature about prognostics models can be found in
were used within a Bayesian framework to update the param- reviews such as statistical data-driven RUL methods [2], ma-
eters of the degradation model (e.g., Paris’ law). The degrada- chinery condition based maintenance approaches[25], [26], and
tion model was initially fed by the results of a stress analysis model selection of RUL estimation by industry[24] and chal-
from a gear dynamic model or finite element model. The dis- lenges in rotating machinery prognostics [27]. Those reviews
tributions of the uncertainty factors were updated via Bayesian have addressed many aspects (e.g., model types, model selec-
inference using the condition monitoring data (simulated crack tion, and challenges, etc.) of the prognostics models. However,
lengths), and an estimation of the remaining useful life based on none of them has been specifically focusing on the hybrid prog-
the degradation model was provided. nostics approaches which leverage the advantages of different
Swanson et al. [19] proposed a Kalman filter method to model prognostics models.
the crack growth in a tensioned steel band. RUL was derived This section focuses on a review of the hybrid prognostics ap-
using the predicted modal frequency state when it reached a pre- proaches for RUL prediction. As the prognostics models are cat-
defined failure threshold. The assumptions for using a Kalman egorized as experience-based models, data-driven models, and
filter are that the system must be linear, and the process and mea- physics-based models in Section I, we review the approaches
surement noise must be white Gaussian and –independent. To using different combinations of the three categories as shown
extend Kalman filters to non-linear systems, variations such as in Fig. 2. According to our findings, the hybrid approaches can
extended Kalman filter and unscented Kalman filter were pro- mainly be categorized into the following five types.
posed. Hu et al. [20] presented a multiscale framework to ad- • H1—Experience-based model + data-driven model
dress different timescales for state of charge (SOC), which is • H2—Experience-based model + physics-based model
quickly time-varying, and state of health (SOH), which changes • H3—Data-driven model + data-driven model
on a larger timescale, for estimation of Lithium-ion battery life. • H4—Data-driven model + physics-based model
The framework used extended Kalman filter approaches, named • H5—Experience-based model + data-driven model +
micro EKF, and macro EKF, to deal with the quick, and large physics-based model
time scales, respectively. Discrete time state space models were
A. H1—Hybrid Approach Incorporating Experience-Based
used with multiple timescales to address the different time in-
Models and Data-Driven Models
tervals for the evolution of SOC and SOH, i.e., to monitor and
predict a system’s health state with time scale separation. Experience-based models incorporate domain knowledge in
Not bound by the linear system or Gaussian noise assumption, the reasoning. Expert systems cannot directly deal with con-
particle filters were proposed for RUL prediction. For example, tinuous variables, and the output of expert systems is often a
Daigle and Goebel [21] overcame the problem of limited sensor discrete event. These reasons have limited the capability of ex-
data by applying a model-based prognostics approach using par- pert systems to derive RUL. Fuzzy-logic systems can deal with
ticle filters with application on solenoid valves. While the actual continuous variables by converting them into fuzzy represen-
intention was the implementation of a computational efficient tations using appropriate designed membership functions, and
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
LIAO AND KÖTTIG: REVIEW OF HYBRID PROGNOSTICS APPROACHES FOR REMAINING USEFUL LIFE PREDICTION OF ENGINEERED SYSTEMS 195
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
196 IEEE TRANSACTIONS ON RELIABILITY, VOL. 63, NO. 1, MARCH 2014
in tension steels bands. Using vibration mode frequencies as two data-driven approaches. The health index of the system
responsive fault features, the Kalman filter performed feature state was calculated by the minimum quantization error (MQE)
tracking and forecasting, and the remaining useful life was es- of the established baseline self-organizing maps trained with
timated using Newton’s method. While the Kalman filter fa- vibration features. The health index was then used as input
cilitated feature trending and prediction, the failure threshold to back propagation neural networks for remaining useful life
strongly depended on the operational conditions of the system, prediction. Yan and Lee [36] estimated the tool wear condition
e.g., the vibrations in a bearing under dynamic load would be (state) in drilling operations via logistic regression analysis
expected to be much higher than that under constant load. To combined with a maximum likelihood technique (for parameter
consider various operational conditions, fuzzy logic was used estimation), based on features extracted from vibration signals
to adapt failure thresholds to the operational conditions. By- using wavelet packet decomposition. Another data-driven
ington et al. [33] presented a methodology for aircraft actuator model, which was an autoregressive moving average (ARMA)
components’ RUL prediction. A fuzzy logic process was im- model, was applied to predict the remaining useful life based
plemented to quantify the level of damage (damage index) of on the estimated tool wear condition.
the system using a predefined set of rules tailored from knowl- There are many other articles, which are not all listed here,
edge of the system and engineering judgment. A physics-based that fall into this hybrid approach category. Both data-driven
method (Kalman filtering) was then applied to predict the pro- health state estimation methods and data-driven prediction
gression of the damage. methods have been extensively studied in the past. It is nat-
With regard to the general framework and interface to inte- ural to use them sequentially for degradation detection and
grate information from different types of models, the hybrid prediction. This type of approach can be easily adopted, and
approaches in the H2 category is similar to the approaches sometimes without demanding data requirements. For example,
in the H1 category. Expert knowledge is mainly used to an unsupervised learning model can be used to build a de-
support the determination of system fault states and catego- scription of the nominal behavior (baseline) of a system using
rization of damage levels; while, instead of using data-driven only data collected during normal operation. The degradation
models, physics-based models are used to perform the actual level can be assessed by continuously calculating a degradation
RUL prediction. The H2 approaches share similar advantages index using the deviation from the established baseline. The
and disadvantages with the H1 approaches, as mentioned in extent of the deviation can be extrapolated to predict the future
Section II-A, from a hybrid approach point of view. Besides, trending using a time series prediction method. RUL can be
the advantages and disadvantages between data-driven models estimated when the predicted deviation exceeds a predefined
and physics-based models have been briefly described in threshold. However, it is not a trivial task to determine the
Sections I-B and I-C. threshold of the degradation index. The degradation index
itself might be a deviation measure estimated from multiple
C. H3—Hybrid Approach Incorporating Multiple Data-Driven sensor measurements. In this case, it is hard to determine the
Models physical meaning of the degradation index, even with domain
Approaches in the H3 category are mainly two-fold. On one knowledge. In addition, without understanding the underline
hand, a data-driven model can be used to estimate the internal (physical) degradation or fault propagation model, relying on
system state (e.g., crack growth rate) when it is not directly mea- the prediction model fitted from a limited amount of data may
surable from the sensor signals. The estimated system state can introduce immeasurable uncertainty to the prediction result.
then be used to extrapolate the future system state to predict With regards to the latter approach, the following literature
RUL using another data-driven model. On the other hand, dif- predicted the system RUL by fusing results from multiple data-
ferent competing data-driven models can be developed for RUL driven models. Gebraeel et al. [13] established the relation-
prediction purposes. The results of different models can be ag- ship between the vibration signals and the bearing operating
gregated to improve the prediction performance by a carefully time using feedforward backpropagation neural networks. Each
designed fusion mechanism. neural network was trained on the data collected from one of
Researchers addressed the first approach of implementing the 25 tested bearings. RUL prediction was accomplished by
a hybrid prognostics approach combining two data-driven weighting the outputs of all 25 neural networks using Weight
models with one estimating and one predicting the system Application to Failure Times (WAFT) and Weight Application
state. Liu et al. [34] presented an approach which combined to Exponential Parameters (WAEP) methods. Peel [37] used an
least squares support vector regression (LSSVR) with hidden ensemble of multilayer perceptron (MLP) and radial basis func-
Markov models (HMMs) for bearing remaining useful life tion (RBF) neural networks for RUL prediction. The neural net-
prediction. Features extracted from vibration sensor signals works provided nonlinear mapping from sensor measurements
were used for offline and online training of the HMMs with to the state space. The neural network ensemble was created
health states representing different health levels during the using a tournament style heuristic. In addition, a Kalman filter
system degradation. LSSVR was used to predict the trend of provided a mechanism to fuse the multiple ensemble outputs
features to infer the future health states through the HMMs and over time, and accomplish RUL prediction. Similarly, Heimes
log-likelihood probabilities. Eventually, the remaining useful [38] applied a recurrent neural network trained by back-propa-
life was estimated as the time until the last health state, e.g., gation method, incorporating an extended Kalman filter to filter
the system failed, of the HMMs was reached. Huang et al. the output. The remaining useful life for the simulated aircraft
[35] predicted the remaining useful life of bearings combining engines was calculated by simply averaging the outputs of the
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
LIAO AND KÖTTIG: REVIEW OF HYBRID PROGNOSTICS APPROACHES FOR REMAINING USEFUL LIFE PREDICTION OF ENGINEERED SYSTEMS 197
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
198 IEEE TRANSACTIONS ON RELIABILITY, VOL. 63, NO. 1, MARCH 2014
prognostics framework that incorporates both data-driven and greatly affected the failure degradation model. Saha et al. [52]
physics-based models, and provided reaming useful life esti- proposed a relevance vector machine (RVM) method to esti-
mates for electronics systems. A data-driven model was first mate the internal parameters of a battery model and the depen-
used to estimate the system state, and detect anomalies by com- dency with time from the extracted features such as voltage, cur-
paring the system conditions with a healthy baseline. Then, ap- rent, power, impedance, frequency, and temperature readings.
propriate physics-based models (e.g., particle filters) were used The RVM method was used to estimate the initial internal bat-
to perform RUL predictions by projecting the future system tery state in the physics-based particle filter method. The RUL
state. Mohanty et al. [48] presented a hybrid model of fatigue was then predicted by propagating the trajectory of the parti-
crack growth in metallic alloys, combining a physics-based state cles. This work was extended in [53] to reduce the uncertainty
space model and a data-driven kernel based Gaussian process in the prediction result using a Rao-Blackwellized Particle Filter
regression model. The state space model was estimated with un- (RBPF) framework. Baraldi et al. [54] presented and evaluated
derlying physics based on the crack closure concept used in the a hybrid approach concerning the accuracy of the remaining
FASTRAN-II model, which incorporated a constraint factor ac- useful life prediction, and the capability of providing an esti-
counting for the stress state on the plastic zone size. The con- mate of the uncertainty. A bagged ensemble of artificial neural
straint factor was calculated with crack length estimations from networks was proposed to model the measurement equation in
the Gaussian process regression. With this constraint factor, the a particle filter scheme, i.e., a data-driven approach was adapted
future crack length was recalculated using the physics-based to build an empirical measurement model (from training data)
state space model. Rosunally et al. [49] used a Bayesian network of a particle filter. While the measurement model in the par-
to predict the remaining useful life of iron structures under cor- ticle filter approach was approximated by the artificial neural
rosion from physics of failure models (linear bilogarithmic law network, the model for the system degradation state evolution
for atmospheric corrosion). A data-driven method using Ma- still has to be known. The authors’ future objective is the sub-
halanobis distance (MD) was used to assess the system health stitution of the system degradation state evolution model with a
status, and detect anomalies based on weight, dimension, and data-driven model, e.g., an ensemble of trained artificial neural
electrical resistance. The RUL distributions from the physics of networks, to use the particle filter scheme in those cases where
failure models and the current health status, which were derived an analytical evolution model of the system is unavailable.
from the distributions of the MD values, were used as inputs Using a data-driven model to establish a mapping from
for a Bayesian network model to obtain updated RUL predic- the measurement to the internal state makes it possible to use
tions. Peng et al. [50] presented an integrated fatigue damage a mathematically sound physics-based model to predict the
diagnostics and prognostics approach. The crack length in fuse- system internal health. Ideally, data collected from the entire
lage lap joints was estimated via lamb wave-based measure- life cycle should be used to establish an exhaustive mapping
ments with piezoelectric actuators and sensors, which were re- (e.g., the vibration root mean square and the corresponding ma-
lated to the crack length through a data-driven multivariate re- terial crack length). Because the physics model is often selected
gression model. The data-driven model was based on signal fea- from a fault initiation or a propagation model, data collected
tures, such as correlation coefficient, phase change, and am- after the detection of an incipient fault (e.g., a small crack) is
plitude, extracted through comparison of the actual received the most effective to build the mapping. In some applications,
signal with a healthy baseline signal. The model parameters of the ground truth of the system health indicator (e.g., the state
the multivariate regression were finally estimated via learning of charge of a lithium-ion battery) is not easy to be measured
from several experiments. Using the estimated crack length and accurately. The date-driven measurement model implicitly
a small time scale model, a Bayes’ theorem based prognostic incorporates uncertainty into the physics-based model, but it
method (maximum relative entropy) was used to incorporate the also avoids exponential error accumulation due to the Paris
physics-based small time scale model into the measurements, type formulation [48].
and thus reduce the uncertainty. Using a sampling technique 2) Use a Data-Driven Model to Replace a System Model in a
(Markov chain Monte Carlo), the crack length was estimated Physics-Based Model: Similar to inferring the internal system
within the Bayesian updating framework resulting in an im- state from measurements by replacing the measurement model
proved state tracking accuracy compared to the deterministic with a data-driven technique as described in Section II-D.1, the
small time scale model. Irving et al. [51] presented a hybrid ap- fault propagation model can also be replaced with a data-driven
proach for crack length forecasting of hydrogen containing pres- model. As expected in [54], the lack of an analytical degrada-
sure vessels. While using a physics-based approach (Paris’ law tion model can be compensated by replacing this model with a
fracture mechanics model) for prediction, historical data were data-driven technique. In this approach, the modeling of physics
employed to identify unacceptable crack forecasts made by a was not explicit in an analytical equation. Instead, the physics of
data-driven approach. Any crack forecast exceeding the min- the degradation progression was considered, and the parameter
imum or maximum crack lengths for each cycle (obtained from estimation process is similar to the framework of the filtering
historical data) was considered as not feasible, and replaced by method (e.g., particle filter). Hence, this approach is still con-
the minimum or maximum values, respectively. In this method, sidered as a hybrid approach in the H4 category. The structure
the data-driven approach was used to estimate the feasibility of of this type of hybrid approach is shown in Fig. 4. In Section I-C,
the estimated system state, which was the crack length. The re- the system degradation or fault progression is modeled with an
sult showed that the hybrid approach did appear to correct the analytical model based on the underlying physics (e.g., Paris’
physics-based model when the change of operating conditions law). However, deriving sophisticated analytical models can be
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
LIAO AND KÖTTIG: REVIEW OF HYBRID PROGNOSTICS APPROACHES FOR REMAINING USEFUL LIFE PREDICTION OF ENGINEERED SYSTEMS 199
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
200 IEEE TRANSACTIONS ON RELIABILITY, VOL. 63, NO. 1, MARCH 2014
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
LIAO AND KÖTTIG: REVIEW OF HYBRID PROGNOSTICS APPROACHES FOR REMAINING USEFUL LIFE PREDICTION OF ENGINEERED SYSTEMS 201
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
202 IEEE TRANSACTIONS ON RELIABILITY, VOL. 63, NO. 1, MARCH 2014
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
LIAO AND KÖTTIG: REVIEW OF HYBRID PROGNOSTICS APPROACHES FOR REMAINING USEFUL LIFE PREDICTION OF ENGINEERED SYSTEMS 203
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
204 IEEE TRANSACTIONS ON RELIABILITY, VOL. 63, NO. 1, MARCH 2014
D. Discussion
While the prediction of measurements via a data-driven Fig. 10. Trajectories of the exponential decay parameter.
predictor in the proposed hybrid prognostics method was
well suited to reduce the sensitivity of the standard particle
filter method to sudden changes, a question arises: should While the system degradation model defined the general prop-
the predicted measurements be considered as more reliable erties of the physics-based degradation process, the measure-
to track the system degradation than the actual physics based ment model evaluated the likelihood of the measurements given
degradation model? The key to this philosophical quandary the internal state, and hence negotiated (potentially) different
can be seen in an increased uncertainty (noise term) in the information about the system’s actual health status. The pur-
measurement equation during the prediction steps. Conse- pose of particle filtering was to identify the latent state of a
quently, the system evolution should be sufficiently robust system described by a Markov process given a set of (noisy)
to not blindly trust and follow the predicted measurements, measurements. However, an inaccurate or even faulty measure-
but rather rely on the degradation model. However, in the ment model can lead to unexpected deviations of the estimated
case of strongly deviating model parameters, e.g., caused system state. This problem occurred in the case study as follows.
by sudden changes in the measurements at the end of the Even when using the ground truth of the internal resistance, the
state tracking phase, the predicted measurements should be measurement model yielded capacity values (which were even-
able to guide the parameters back to reasonable values. The tually compared with the capacity measurements within the par-
implication of the predicted measurements therefore should ticle filter) that deviated from the measured capacity values.
be regarded as failure avoidance, not as measurements that This difference was caused by the training methodology and
are unconditionally considered as trustworthy. As shown in the character of the SVR model. Within this nonlinear regres-
Fig. 10, the system degradation model implied a degradation sion method, support vectors were created from a set of training
process of exponential type, i.e., . data that should cover the scope of the batteries’ internal resis-
This implication hardly held true for the capacity degradation tances and the corresponding capacities sufficiently. However,
process of the testing battery. Moreover, there were numerous the training data can never reflect the entire range of possible
sudden changes or spikes in the measured capacity values that degradation processes, and thus a discrepancy between the SVR
considerably differed from the assumed exponential model. The estimation and the real capacity value was introduced. Further-
exponential decay parameter was expected to be rather con- more, the SVR model with a radial basis function kernel created
stant, but the vector retrieved from the actual measurements a smooth measurement model from a (very likely) not smooth
( ) was significantly varying, set of training data, which introduced a certain deviation as well.
and manifested in the way that the real degradation process The presented case study showed that a carefully designed
significantly differed from the assumed model. Although the hybrid prognostics approach can outperform the widely used
determination of the corresponding uncertainties can be a standard particle filter approach (physics-based prognostics) in
balancing act, the superior performance of the proposed hybrid terms of both the prediction accuracy and the corresponding
prognostics method over the standard particle filter suggested uncertainty. The particle filter as de facto state-of-the-art in
that the proposed method was reasonable, and can improve physics-based prognostics lacked the robustness to manage
the prognostic performance, especially in those cases where very noisy measurements, or even sudden, but not continuous,
the measurements were very noisy and non-monotonic, or the changes in the health state of the device that falsified the
actual system degradation process significantly differed from estimated system state during the state tracking phase. Conse-
the physics-based degradation model, or both. However, the quently, the prediction performance was prone to suffer from
predicted measurements may not be considered as a panacea to a wrong identification of the system parameters. The applica-
overcome limited knowledge of the underlying physics of the tion of a data-driven prediction method within the presented
degradation process. hybrid prognostics method alleviated this risk, and greatly
The prognostics performance of the proposed hybrid prog- improves the overall prediction performance. The case study
nostics method significantly depended on the performance of on RUL prediction of lithium-ion batteries prompted that the
both the measurement model and the system degradation model. presented hybrid prognostics method can be used under limited
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
LIAO AND KÖTTIG: REVIEW OF HYBRID PROGNOSTICS APPROACHES FOR REMAINING USEFUL LIFE PREDICTION OF ENGINEERED SYSTEMS 205
knowledge of the actual degradation process, and a non-distinct model without exhaustive knowledge or extensive experi-
degradation behavior. ment when the system is complex. In the case as described
in [51], the adopted physics-based model may not always
match the actual system degradation when the change of
IV. CONCLUSION
operating condition greatly affects the fault propagation
Prognostics models were introduced, and mainly catego- model. It is also critical to determine the failure threshold
rized as experience-based models, data-driven models, and in RUL prediction applications. Expert knowledge or rules
physics-based models. The concepts of those models were can help with the task, while most knowledge is subjective
introduced, followed by typical practices in the literature, and information which may have issues in robustness, consis-
the advantages and disadvantages were briefly summarized. tency, and maintenance. It is essential to thoroughly inves-
A review specifically focusing on the hybrid prognostics ap- tigate the data availability, existing knowledge, and system
proaches, which leverage the advantages of different models, information to systematically determine the appropriate
was given according to different combinations of the aforemen- type of a hybrid approach. It may also be necessary to lay
tioned three types of prognostics models. out an additional data collection plan and take into consid-
Most of the PHM literature concentrated on the hybrid eration the constraints from a business point of view (e.g.,
approach of combining data-driven models and physics-based budget and deadline).
models, as described in Section II-D. In this type of hybrid • The second challenge lies in designing the appropriate
approach, the physics-based model (e.g., particle filter) in- fusing mechanism for both heterogeneous information
corporates first principles (e.g., Paris’ law of crack growth), and competing models. Similar as experienced in [28],
and the model parameters are identified and updated with data has different formats such as continuous wave forms,
measurement data. RUL is derived by projecting the system discrete samples, if-then rules, voice recordings, text
internal state estimation to the future until a failure threshold logs, etc. Data can also be collected at different sampling
is reached. The data-driven model in this hybrid approach has rates, saved in different media, and collected with missing
been used for anomaly detection to trigger RUL prediction values. Hence, an information fusing mechanism should
process, estimation of the internal system state from measure- be well-defined to consolidate and make use of all kinds of
ment data, and as a replacement of a system degradation model data. The hybrid approach is the combination of different
in the physics-based model prediction. The hybrid prognostics models. Some competing models may generate different
models of combining data-driven models and physic-based prediction results. The answer of the result from which
models also include approaches of fusing the results obtained model should be considered as the ultimate prediction and
from separate models to improve the RUL estimation. A hybrid whether or how to aggregate the results from different
prognostics method was proposed, and a set of battery data models as the final prediction may vary case by case. The
was used as a case study to demonstrate the effectiveness of fusing mechanism has been successfully applied within
combining the physics-based models and data-driven models. the context of classification with ensemble methods (e.g.,
Sparse literature has mentioned the practices of using the hy- [70]) in which the output of multiple classifiers were used
brid approach which incorporates all of the experience-based to derive better result than any single classifier. Either
models, data-driven models, and physics-based models. It is heuristic or theoretically sound fusing mechanisms need
potentially beneficial to fuse all types of information (e.g., do- to be designed within the context of RUL prediction.
main knowledge, maintenance feedback, and condition data and • The third challenge is to manage the prediction uncertain-
physics), and to leverage the strengths of all types of models to ties in RUL prediction. Robust prediction under model
better manage the prediction uncertainty. Despite the opportuni- uncertainty is a main requirement for hybrid prognostics
ties brought by the hybrid approach, challenges remain in how approaches. The representation of the uncertainty for
to design the fusing method (e.g., using neural network to in- prognostics is a difficult task because prognostics involves
corporate rule based experience and condition data [28]) to inte- both subjective and objective uncertainties, and operates
grate heterogeneous information, how to aggregate results from over the time horizon from the past, through the present,
different competing models (e.g., using Dempster-Shafer-based and into the future [71]. For a single prediction model,
regression [59] or Bayesian framework [63]), and how to utilize it is possible to model the prediction in a mathemati-
data-driven models to reduce the prediction uncertainty (e.g., cally sound manor. For example, for data-driven models,
using data-driven models to estimate the measurement model the model prediction uncertainty for an auto-regressive
[54] or replace a system model [55]). We summarize the future moving average model, and for a neural network model,
challenges of a hybrid prognostics approach as follows. can be modeled as described in[72], and [73], respectively.
• The first challenge is how to determine the right type of For physics-based models, three methods were proposed
hybrid approach depending on the available data and in- in [57] to model prediction uncertainty of long-term pre-
formation. It is ideal to develop a RUL prediction model diction using the particle filter method. When combining
based on few data or no data, especially for newly commis- multiple prediction models, especially in different cate-
sioned systems. However, the data quantity and complete- gories, the uncertainty management becomes a non-trivial
ness would not be sufficient for many data-driven models. issue, not mentioning how the information and model fu-
Physics-based models can shed some light in this case, sion mechanism are designed. Sparse literature mentioned
while it is generally hard to obtain an accurate physics how to manage the uncertainty of a hybrid approach when
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
206 IEEE TRANSACTIONS ON RELIABILITY, VOL. 63, NO. 1, MARCH 2014
combining multiple models specifically for RUL predic- [18] F. Zhao, Z. Tian, and Y. Zeng, “Uncertainty quantification in gear
tion. A Bayesian statistical approach was proposed in [74] remaining useful life prediction through an integrated prognostics
method,” IEEE Trans. Rel., vol. 62, no. 1, pp. 146–159, Mar. 2012.
to handle various sources of uncertainty with prediction [19] D. C. Swanson, J. M. Spencer, and S. H. Arzoumanian, “Prognostic
models of a relevance vector machine and a particle modelling of crack growth in a tensioned steel band,” Mechan. Syst.
filter. A method was described in [60] to illustrate how Signal Process., vol. 14, no. 5, pp. 789–803, 2000.
[20] C. Hu, B. Youn, T. Kim, and J. Chung, “Online estimation of
the fusion of two different prognostics approaches can lithium-ion battery state-of-charge and capacity with a multiscale
produce a result that is more accurate and generate more filtering technique,” in Proc. Annu. Conf. Prognost. andHealth Man-
narrow uncertainty bounds than either approach alone. agement Soc., 2011, vol. 5.
[21] M. Daigle and K. Goebel, “Improving Computational Efficiency of
A systematic methodology still needs to be developed to Prediction in Model-Based Prognostics using the Unscented Trans-
manage the uncertainties of all types of hybrid prognostics form,” Tech. Rep. DTIC Document, 2010.
approaches so that the probability as a degree of belief that [22] M. Daigle and K. Goebel, “Model-based prognostics under limited
sensing,” in Proc. IEEE Aerospace Conf., 2010, pp. 1–12.
a prediction is true can be determined. [23] J. Luo, M. Namburu, K. Pattipati, L. Qiao, M. Kawamoto, and S. Chi-
gusa, “Model-based prognostic techniques,” in Proc, IEEE Syst. Readi-
ness Technol. Conf., 2003, pp. 330–340.
REFERENCES [24] J. Sikorska, M. Hodkiewicz, and L. Ma, “Prognostic modelling op-
[1] ISO 13381-1, Condition Monitoring and Diagnostics of Ma- tions for remaining useful life estimation by industry,” Mechanical
chines—Prognostics—Part 1: General Guidelines: International Syst. Signal Process., vol. 25, no. 5, pp. 1803–1836, 2011.
Standards Organization International Standards Organization Std, [25] A. Jardine, D. Lin, and D. Banjevic, “A review on machinery diag-
2004. nostics and prognostics implementing condition-based maintenance,”
[2] X. Si, W. Wang, C. Hu, and D. Zhou, “Remaining useful life estima- Mechanical Syst. Signal Process., vol. 20, no. 7, pp. 1483–1510, 2006.
tion—A review on the statistical data driven approaches,” Eur. J. Op- [26] R. Kothamasu, S. Huang, and W. VerDuin, “System health monitoring
erat. Res., vol. 213, no. 1, pp. 1–14, 2011. and prognostics—A review of current paradigms and practices,” Int. J.
[3] T. Biagetti and E. Sciubba, “Automatic diagnostics and prognostics of Adv. Manufacturing Technol., vol. 28, no. 9, pp. 1012–1024, 2006.
energy conversion processes via knowledge-based systems,” Energy, [27] A. Heng, S. Zhang, A. Tan, and J. Mathew, “Rotating machinery prog-
vol. 29, no. 12, pp. 2553–2572, 2004. nostics: State of the art, challenges and opportunities,” Mechan. Syst.
[4] A. Majidian and M. Saidi, “Comparison of fuzzy logic and neural net- Signal Process., vol. 23, no. 3, pp. 724–739, 2009.
work in life prediction of boiler tubes,” Int. J. Fatigue, vol. 29, no. 3, [28] A. Garga, K. McClintic, R. Campbell, C. Yang, M. Lebold, T. Hay,
pp. 489–498, 2007. and C. Byington, “Hybrid reasoning for prognostic learning in cbm
[5] K. Medjaher, D. Tobon-Mejia, and N. Zerhouni, “Remaining useful life systems,” in Proc. IEE Aerospace Conf., 2001, vol. 6, pp. 2957–2969.
estimation of critical components with application to bearings,” IEEE [29] B. Satish and N. Sarma, “A fuzzy BP approach for diagnosis and
Trans. Rel., vol. 61, no. 2, pp. 292–302, Jun. 2012. prognosis of bearing faults in induction motors,” in Proc IEEE Power
[6] Y. Peng and M. Dong, “A hybrid approach of HMM and grey model Engin. Soc. General Meet., 2005, pp. 2291–2294.
for age-dependent health prediction of engineering assets,” Expert Syst. [30] R. Chinnam and P. Baruah, “A neuro-fuzzy approach for estimating
Appl., vol. 38, no. 10, pp. 12 946–12 953, 2011. mean residual life in condition-based maintenance systems,” Int. J. Ma-
[7] D. Liu, Y. Luo, Y. Peng, X. Peng, and M. Pecht, “Lithium-ion battery terials Product Technol., vol. 20, no. 1, pp. 166–179, 2004.
remaining useful life estimation based on nonlinear AR model com- [31] G. Zhang, S. Lee, N. Propes, Y. Zhao, G. Vachtsevanos, A. Thakker,
bined with degradation feature,” in Proc. Annu. Conf. Prognost. Health and T. Galie, “A novel architecture for an integrated fault diagnostic/
Management Soc., 2012, vol. 3. prognostic system,” in Proc. AAAI Symp., Stanford, California, 2002.
[8] J. Yan, M. Koc, and J. Lee, “A prognostic algorithm for machine per- [32] D. C. Swanson, “A general prognostic tracking algorithm for predic-
formance assessment and its application,” Product. Planning Contr., tive maintenance,” in Proc. IEEE Aerospace Conf., 2001, vol. 6, pp.
vol. 15, no. 8, pp. 796–801, 2004. 2971–2977.
[9] B. Lamoureux, J. Massé, and N. Mechbal, “An approach to the health [33] C. S. Byington, M. Watson, and D. Edwards, “Data-driven neural
monitoring of a pumping unit in an aircraft engine fuel system,” in network methodology to remaining life predictions for aircraft actu-
Proc. 1st Eur. Conf. Prognost. Health Management Soc., 2012, vol. 3. ator components,” in Proc. IEEE Aerospace Conf., 2004, vol. 6, pp.
[10] H. Guo, S. Watson, P. Tavner, and J. Xiang, “Reliability analysis for 3581–3589.
wind turbines with incomplete failure data collected from after the [34] Z. Liu, Q. Li, and C. Mu, “A hybrid LSSVR-HMM based prognos-
date of initial installation,” Rel. Eng. Syst. Safety, vol. 94, no. 6, pp. tics approach,” in Proc. 4th Int. Conf. Intell. Human-Machine Syst. Cy-
1057–1063, 2009. bernet., 2012, vol. 2, pp. 275–278.
[11] H. Sutherland, T. Repoff, M. House, and G. Flickinger, “Prognostics, A [35] R. Huang, L. Xi, X. Li, C. Richard Liu, H. Qiu, and J. Lee, “Residual
new look at statistical life prediction for condition-based maintenance,” life predictions for ball bearings based on self-organizing map and back
in Proc. IEEE Aerospace Conf., 2003, vol. 7, pp. 3131–3136. propagation neural network methods,” Mechan. Syst. Signal Process.,
[12] K. Goode, J. Moore, and B. Roylance, “Plant machinery working life vol. 21, no. 1, pp. 193–207, 2007.
prediction method utilizing reliability and condition-monitoring data,” [36] J. Yan and J. Lee, “A hybrid method for on-line performance assess-
in Proc. Institut. Mechan. Engineers, Part E: J. Process Mechanical ment and life prediction in drilling operations,” in Proc. IEEE Int. Conf.
Eng., 2000, vol. 214, pp. 109–122, 2. Automat. Logistics, 2007, pp. 2500–2505.
[13] N. Gebraeel, M. Lawley, R. Liu, and V. Parmeshwaran, “Residual [37] L. Peel, “Data driven prognostics using a Kalman filter ensemble of
life predictions from vibration-based degradation signals: A neural neural network models,” in Proc. Int. Conf. Prognost. Health Manage-
network approach,” IEEE Trans. Ind. Electron., vol. 51, no. 3, pp. ment, 2008, pp. 1–6.
694–700, 2004. [38] F. Heimes, “Recurrent neural networks for remaining useful life esti-
[14] A. Heng, A. Tan, J. Mathew, N. M. D. Banjevic, and A. Jardine, “In- mation,” in Proc. Int. Conf. Prognost. Health Management, 2008, pp.
telligent condition-based prediction of machinery reliability,” Mechan. 1–6.
Syst. Signal Process., vol. 23, no. 5, pp. 1600–1614, 2009. [39] S. Du, J. Lv, and L. Xi, “Degradation process prediction for rotational
[15] Z. He, S. Wang, K. Wang, and K. Li, “Prognostic analysis based on machinery based on hybrid intelligent model,” Robot. Comput.-Integr.
hybrid prediction method for axial piston pump,” in Proc. 10th IEEE Manufacturing, vol. 28, no. 2, pp. 190–207, 2012.
Int. Conf. Ind. Inf. (INDIN 2012), 2012, pp. 688–692. [40] J. Celaya, A. Saxena, S. Saha, and K. Goebel, “Prognostics of power
[16] J. Liu, D. Djurdjanovic, J. Ni, N. Casoetto, and J. Lee, “Similarity based mosfets under thermal stress accelerated aging using data-driven and
method for manufacturing process performance prediction and diag- model-based methodologies,” in Proc. Int. Conf. Prognost. Health
nosis,” Comput. Ind., vol. 58, no. 6, pp. 558–566, 2007. Management, 2011.
[17] T. Wang, J. Yu, D. Siegel, and J. Lee, “A similarity-based prognostics [41] M. Orchard and G. Vachtsevanos, “A particle filtering approach for
approach for remaining useful life estimation of engineered systems,” on-line failure prognosis in a planetary carrier plate,” Int. J. Fuzzy
in Proc. Int. Conf. Prognost. Health Manag., 2008, pp. 1–6. Logic Intell. Syst., vol. 7, no. 4, pp. 221–227, 2007.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.
LIAO AND KÖTTIG: REVIEW OF HYBRID PROGNOSTICS APPROACHES FOR REMAINING USEFUL LIFE PREDICTION OF ENGINEERED SYSTEMS 207
[42] S. Kumar, M. Torres, Y. Chan, and M. Pecht, “A hybrid prognostics [62] J. Xu and L. Xu, “Health management based on fusion prognostics for
methodology for electronic products,” in Proc. IEEE Int. Joint Conf. avionics systems,” J. Syst. Engin. Electron., vol. 22, no. 3, pp. 428–436,
Neural Netw., 2008, pp. 3479–3485. 2011.
[43] H. Zhang, R. Kang, and M. Pecht, “A hybrid prognostics and health [63] R. Orsagh, J. Sheldon, and C. Klenke, “Prognostics/diagnostics for gas
management approach for condition-based maintenance,” in IEEE Int. turbine engine bearings,” in Proc. IEEE Aerospace Conf., 2003, pp.
Conf. on Industrial Engin. and Engin. Management. IEEM 2009, 2009, 1165–1173.
pp. 1165–1169. [64] G. Gola and B. Nystad, “From measurement collection to remaining
[44] N. Patil, D. Das, C. Yin, H. Lu, C. Bailey, and M. Pecht, “A fusion useful life estimation: Defining a diagnostic-prognostic frame for op-
approach to igbt power module prognostics,” in Proc. 10th Int. Conf. timal maintenance scheduling of choke valves undergoing erosion,”
Thermal, Mechanical Multi-Phys. Simulat. Experiments Microelec- in Proc. Annu. Conf. Prognost. Health Management Soc., 2011, pp.
tronics Microsyst., 2009, pp. 1–5. 26–29.
[45] S. Cheng and M. Pecht, “A fusion prognostics method for remaining [65] K. Goebel, B. Saha, A. Saxena, J. Celaya, and J. Christophersen,
useful life prediction of electronic products,” in Proc. IEEE Int. Conf. “Prognostics in battery health management,” IEEE Instrumen.
Automat. Sci. Eng. , 2009, pp. 102–107. Measur. Mag., vol. 11, no. 4, pp. 33–40, 2008.
[46] M. Pecht and R. Jaai, “A prognostics and health management roadmap [66] B. Saha and K. Goebel, “Battery data set,” in NASA Ames Prognostics
for information and electronics-rich systems,” Microelectronics Reli- Data Repository. Moffett Field, CA: NASA Ames, 2007.
abil., vol. 50, no. 3, pp. 317–323, 2010. [67] R. Douc and O. Cappé, “Comparison of resampling schemes for par-
[47] Y. Xing, N. Williard, K. Tsui, and M. Pecht, “A comparative review of ticle filtering,” in Proc. 4th Int. Symp. Image Signal Process. Anal.,
prognostics-based reliability methods for lithium batteries,” in Proc. 2005, pp. 64–69.
Prognost. System Health Management Conf., 2011, pp. 1–6. [68] W. Luo, C. Lv, L. Wang, and C. Liu, “Study on impedance model
[48] S. Mohanty, R. Teale, A. Chattopadhyay, P. Peralta, and C. Willhauck, of li-ion battery,” in Proc. 2011 6th IEEE Conf. Ind. Electron. Appl.
“Mixed Gaussian process and state-space approach for fatigue crack (ICIEA), , 2011, pp. 1943–1947.
growth prediction,” in Proc. Int. Workshop Structural Heath Moni- [69] A. Saxena, J. Celaya, E. Balaban, K. Goebel, B. Saha, S. Saha, and
toring, 2007, vol. 2, pp. 1108–1115. M. Schwabacher, “Metrics for evaluating performance of prognostic
[49] Y. Rosunally, S. Stoyanov, C. Bailey, P. Mason, S. Campbell, and G. techniques,” in Proc. Int. Conf. Prognost. Health Manag., 2008, pp.
Monger, “Prognostics framework for remaining life prediction of cutty 1–17.
sark iron structures,” in Proc. Annu. Conf. Prognost. Health Manage- [70] T. G. Dietterich, “Ensemble methods in machine learning,” in Multiple
ment Soc., 2009. Classifier Syst.. : Springer, 2000, pp. 1–15.
[50] T. Peng, J. He, Y. Liu, A. Saxena, J. Celaya, and K. Goebel, “Integrated [71] P. Wang and G. Vachtsevanos, “Fault prognostics using dynamic
fatigue damage diagnosis and prognosis under uncertainties,” in Proc. wavelet neural networks,” AI EDAM, vol. 15, no. 04, pp. 349–365,
Annu. Conf. Prognost. Health Management Soc., 2012, vol. 3. 2001.
[51] P. Irving, O. Eker, and F. Camci, “Prognostics control of cracking in [72] S. M. Pandit and S.-M. Wu, Time Series and System Analysis with Ap-
structures and components operating in hydrogen environments,” in plications. : RE Krieger Publishing Company, 1990.
Proc. Int. Hydrogen Conf., 2012. [73] G. Yu, H. Qiu, D. Djurdjanovic, and J. Lee, “Feature signature pre-
[52] B. Saha, K. Goebel, S. Poll, and J. Christophersen, “An integrated ap- diction of a boring process using neural network modeling with confi-
proach to battery health monitoring using Bayesian regression and state dence bounds,” Int. J. Adv. Manufacturing Technol., vol. 30, no. 7-8,
estimation,” in Proc IEEE Autotestcon, 2007, pp. 646–653. pp. 614–621, 2006.
[53] B. Saha, K. Goebel, S. Poll, and J. Christophersen, “Prognostics [74] B. Saha and K. Goebel, “Uncertainty management for diagnostics and
methods for battery health monitoring using a Bayesian framework,” prognostics of batteries using Bayesian techniques,” in Proc. IEEE
IEEE Trans. Instrumentation Measur., vol. 58, no. 2, pp. 291–296, Aerospace Conf., 2008, pp. 1–8.
2009.
[54] P. Baraldi, M. Compare, S. Sauco, and E. Zio, “Fatigue crack growth
prognostics by particle filtering and ensemble neural networks,” in
Proc 1st Eur. Conf. Prognost.Health Management Soc., 2012, vol. 3.
[55] C. Chen, G. Vachtsevanos, and M. Orchard, “Machine remaining useful
life prediction: An integrated adaptive neuro-fuzzy and high-order par-
ticle filtering approach,” Mechanical Syst. and Signal Process., vol. 28,
pp. 597–607, 2011. Linxia Liao is a member research staff in the intelligent systems laboratory at
[56] C. Chen, B. Zhang, G. Vachtsevanos, and M. Orchard, “Machine condi- Palo Alto Research Center (PARC, a Xerox Company). He was a research sci-
tion prediction based on adaptive neuro-fuzzy and high-order particle entist at Siemens Corporation, Corporate Technology in the US. He received
filtering,” IEEE Trans. Ind. Electron., vol. 58, no. 9, pp. 4353–4364, his Ph.D. degree in Industrial Engineering from the University of Cincinnati.
2011. He received his M.S. degree in Mechanical and Electronic Engineering from
[57] M. E. Orchard, A Particle Filtering-Based Framework for On-Line Huazhong University of Science and Technology (HUST, China), and B.S. de-
Fault Diagnosis and Failure Prognosis, 2007. gree in Mechanical Manufacture and Automation from HUST. His interests in-
[58] J. Liu, W. Wang, F. Ma, Y. Yang, and C. Yang, “A data-model-fusion clude predictive and prescriptive analytics, system and components fault diag-
prognostic framework for dynamic system state forecasting,” Engin. nostics and prognostics, signal processing, and machine learning algorithms as
Appl. Artif. Intell., vol. 25, pp. 814–823, 2012. well as their integration on embedded systems.
[59] K. Goebel, N. Eklund, and P. Bonanni, “Fusing competing prediction
algorithms for prognostics,” in Proc. IEEE Aerospace Conf., 2006, p.
10.
[60] K. Goebel, N. Eklund, and P. Bonanni, Prognostic Fusion for Uncer- Felix Köttig is a masters student at Dresden University of Technology (TUD),
tainty Reduction. Wright-Patterson AFB, OH, USA: Defense Tech- Germany, and a research intern at Siemens Corporation, Corporate Technology
nical Information Center, 2007. US. He is supported by the German national scholarship program, and a student
[61] G. Bartram and S. Mahadevan, “Prognostics and health monitoring assistantship at the department of Clinical Sensoring and Monitoring (TUD).
in the presence of heterogeneous information,” in Proc. Annu. Conf. His research interests include data analysis, predictive modeling, and medical
Prognost. Health Management Soc., 2012, vol. 3. imaging (optical coherence tomography).
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2021 at 10:42:54 UTC from IEEE Xplore. Restrictions apply.